Datasets:

bskang
/

CVPR2023_title_abstract_intro

Dataset card Viewer Files Files and versions Community

Split (1)

train · 2.34k rows

title stringlengths 28 135	abstract stringlengths 0 12k	introduction stringlengths 0 12k
Wang_ProTeGe_Untrimmed_Pretraining_for_Video_Temporal_Grounding_by_Video_Temporal_CVPR_2023	Abstract Video temporal grounding (VTG) is the task of localiz- ing a given natural language text query in an arbitrarily long untrimmed video. While the task involves untrimmed videos, all existing VTG methods leverage features from video backbones pretrained on trimmed videos. This is largely due to the lack of large-scale well-annotated VTG dataset to perform pretraining. As a result, the pretrained features lack a notion of temporal boundaries leading to the video-text alignment being less distinguishable between correct and incorrect locations. We present ProT ´eG´e as the first method to perform VTG-based untrimmed pretrain- ing to bridge the gap between trimmed pretrained back- bones and downstream VTG tasks. ProT ´eG´e reconfigures the HowTo100M dataset, with noisily correlated video-text pairs, into a VTG dataset and introduces a novel Video-Text Similarity-based Grounding Module and a pretraining ob- jective to make pretraining robust to noise in HowTo100M. Extensive experiments on multiple datasets across down- stream tasks with all variations of supervision validate that pretrained features from ProT ´eG´e can significantly outper- form features from trimmed pretrained backbones on VTG.	1. Introduction Video temporal grounding (VTG) is the video-language multimodal task of localizing which part of an arbitrarily long untrimmed video can be best associated with a given natural language text query. VTG has a wide range of ap- plications, such as information retrieval and robotics. Fig- ure 1 shows a sample video-text pair for the VTG task and illustrates the primary challenge in grounding an uncon- strained natural language text query in a long untrimmed video, namely, the need for a fine-grained understanding of the spatio-temporal dynamics in the video. ⋆Authors with equal contribution. This work was done as Lan Wang’s internship project at Microsoft. Downstream Untrimmed VideoGround Truth“A person kneeling on the floor talks on a phone.” (Q1)“The person pours something into a glass.(Q2)0.0-8.4s13.3-22.4sVideo-textPretraining on TrimmedVideosProTéGé: Video-text Pretraining on UntrimmedVideos0.0-10.4s13.4-19.3s0.0-8.9s13.4-22.3s……… Cosine Similarity of Downstream UntrimmedVideo Features with Q2“Use a watercolor pen to fill in the gaps.” ……“Next step is to fill the channel”(a) (b) Video-text Pretraining on TrimmedVideos(c) ProTéGé: Video-text Pretraining on UntrimmedVideos “Reinforce the handrail with screws.”Cosine Similarity of Downstream UntrimmedVideo Features with Q2 “The person climbing mountain.”Figure 1. (a) Comparison between video-text trimmed and untrimmed pretraining on grounding text Q1 and Q2 in an untrimmed video. Untrimmed video-text pretraining shows stronger grounding capability. (b) and (c) show box plots of co- sine similarity after joint video-text pretraining on trimmed and untrimmed videos respectively, between video features aligning with text (blue) and not aligning with text (red). We observe that compared to using trimmed videos (b), cosine similarities of video features aligning with text are higher and farther apart from that of video features not aligning with text when using untrimmed videos (c), thus illustrating the impact of untrimmed pretraining. While there are multiple approaches for VTG, all exist- ing methods, to the best of our knowledge, rely on video backbones pretrained on trimmed videos (such as Kinetics [15]) to obtain the visual features as part of their respective approaches. Such a design choice introduces a disconnect between the downstream VTG task on untrimmed videos and the trimmed videos used for pretraining the model from which video features are derived. For example, Fig 1 shows that the grounding predictions (in orange), when using a backbone jointly pretrained on trimmed videos and text, do not match adequately with the ground truth. Due to pretraining on trimmed videos, the video backbone is in- sensitive to temporal boundaries since the training objec- tive is to associate an entire trimmed video to a label/text This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6575 query [43, 44]. The backbone, therefore, does not have an explicit ability to localize, i.e., associate the given query to only the most relevant part of the long untrimmed video. As a result, the cosine similarity between video features align- ing and not aligning with the text query are indistinguish- able, as shown in Fig 1b. Inspired by the advantage shown in other tasks where pretraining and downstream setup match [3, 7, 23, 26], we hypothesize that formulating the pretraining itself as a VTG task on untrimmed videos can improve downstream ground- ing performance. The untrimmed pretraining will equip the model with a more accurate and fine-grained understand- ing of temporal boundaries within a given untrimmed video (as evidenced by the more precise predictions in blue in Fig. 1). We introduce ProT ´eG´e, Untrimmed Pretraining for Video Te mporal Grounding by Video T emporal Ground- ing. ProT ´eG´e is the first approach to formulate pretraining as a VTG task to bridge the gap between video backbones pretrained on trimmed videos and downstream VTG tasks working with untrimmed videos. A critical challenge impeding this untrimmed pretrain- ing is the scarcity of large-scale well-annotated video grounding datasets. There are, however, datasets such as HowTo100M [30] and Youtube-8M [1], with over a million untrimmed videos and corresponding subtitled text gener- ated via automated speech-to-text APIs. One can poten- tially employ them for untrimmed pretraining as a VTG task. However, as noted by prior methods [12,29,41], since the text is derived from subtitles, the video regions are only noisily-correlated with the subtitled text, rendering the util- ity of these video-text pairs for grounding a non-trivial task. To overcome the aforementioned challenges in leverag- ing large-scale untrimmed video datasets, we first propose a novel approach to transform them into VTG datasets. Then we introduce a novel video-text similarity grounding mod- ule along with an optimization objective that allows the pre- training to be robust to the noisy video-text correlations present in these datasets. In this work, we use ProT ´eG´e with HowTo100M in par- ticular. To transform HowTo100M into a VTG dataset, ProT ´eG´e introduces aggregated subtitles to concatenate one or more subtitles to form the text query and randomly sam- ples an untrimmed video segment around the query. Ag- gregated subtitles allow ProT ´eG´e to incorporate arbitrarily long text queries larger than the average 4s duration of a sin- gle subtitle. This way, we can synthesize millions of video- text grounding pairs for VTG pretraining. Using these pairs, ProT ´eG´e performs pretraining with our novel Video-Text Similarity-based Grounding Module (VT-SGM). VT-SGM creates a 2D-proposal grid by computing the cosine similar- ity between the text query and the different temporal regions of the untrimmed video. It then learns to maximize the sim- ilarity between the query and the part that is most relevantto it. This is achieved via our novel pretraining objective that incorporates a distance-based localization loss which uses the noisy ground truth and a combination of inter-video and intra-video alignment losses. This allows the objective to balance the training via the noisy ground truth and mul- timodal video-text representation learning. We show that ProT ´eG´e is very effective for VTG as a downstream task. It significantly outperforms backbones pretrained on trimmed videos on standard datasets across all variations of supervi- sion. We summarize our contributions as, 1. We propose ProT ´eG´e, the first pretraining method for- mulated as a video temporal grounding task to bridge the gap between pretraining and downstream video temporal grounding tasks in untrimmed videos. 2. We propose a novel algorithm including aggregated subtitles , a Video-Text Similarity-based Grounding Module, and a pretraining objective to leverage large- scale untrimmed video dataset HowTo100M with noisy video-text pairs. 3. Extensive experiments on standard datasets across multiple downstream tasks with different levels of su- pervision validate that our approach significantly im- proves the performance across all benchmarks.
Wang_Deep_Hashing_With_Minimal-Distance-Separated_Hash_Centers_CVPR_2023	Abstract Deep hashing is an appealing approach for large-scale image retrieval. Most existing supervised deep hashing methods learn hash functions using pairwise or triple image similarities in randomly sampled mini-batches. They suffer from low training efficiency, insufficient coverage of data distribution, and pair imbalance problems. Recently, cen- tral similarity quantization (CSQ) attacks the above prob- lems by using “hash centers” as a global similarity metric, which encourages the hash codes of similar images to ap- proach their common hash center and distance themselves from other hash centers. Although achieving SOTA retrieval performance, CSQ falls short of a worst-case guarantee on the minimal distance between its constructed hash centers, i.e. the hash centers can be arbitrarily close. This pa- per presents an optimization method that finds hash cen- ters with a constraint on the minimal distance between any pair of hash centers, which is non-trivial due to the non- convex nature of the problem. More importantly, we adopt the Gilbert-Varshamov bound from coding theory, which helps us to obtain a large minimal distance while ensuring the empirical feasibility of our optimization approach. With these clearly-separated hash centers, each is assigned to one image class, we propose several effective loss functions to train deep hashing networks. Extensive experiments on three datasets for image retrieval demonstrate that the pro- posed method achieves superior retrieval performance over the state-of-the-art deep hashing methods.	1. Introduction Hashing methods are widely-used in large-scale image retrieval due to their excellent efficiency in both storage and retrieval. Recently, much effort has been devoted to deep-learning-based hashing ( deep hashing ) methods for image retrieval. They use deep neural networks to learn hash functions that encode similar/dissimilar images to nearby/faraway binary codes, respectively. Most of the ex- *Corresponding Authoristing deep hashing methods train models on pairwise/triple similarities among training samples in randomly sampled mini-batches (e.g., [1, 10, 18, 22, 24]). Very recently, Yuan et al. [26] pointed out that these methods lead to restricted performance due to three problems: low-efficiency to obtain global similarity of the dataset, incomplete coverage of data distribution that harms the discriminability of the generated hash codes, and ineffectiveness on imbalanced amount of similar/dissimilar data pairs. They then proposed central similarities that finds mutually separated hash centers for each class of similar images, and uses these centers to en- sure small distances between the hash codes of similar im- ages and large distances between those of dissimilar ones. For deep hashing methods that use hash centers, it is cru- cial to construct well-separated hash centers, i.e. the Ham- ming distance between two hash centers should be signifi- cantly larger than the Hamming distance between the hash codes of two similar images, which makes it challenging to generalize to various length of hash code and different number of image classes. For instance, CSQ [26] adopts Hadamard matrix and Bernoulli sampling to produce hash centers with nice properties that any two centers’ Hamming distance is on average half of the hash code length. How- ever, pairs of hash centers constructed in this way can be arbitrarily small in the worst case, i.e. zero in Hamming distance (see Table 4). These degenerated hash centers are expected to harm the retrieval performance. To address this issue, we propose a novel deep hashing method that uses an optimization procedure to produce hash centers, with an additional constraint on a given minimal distance dbetween any pair of hash centers. The value of d is derived using the Gilbert-Varshamov bound [20] adopted from coding theory, which help us to find a large dwhile ensuring the feasibility of our optimization procedure. As shown in Fig.1, the proposed method employs a two- stage pipeline. In Stage 1, we tackle the optimization prob- lem stated above to produce clearly-separated hash centers. To solve this optimization problem, we propose an alternat- ing optimization procedure that relies on the ℓp-box binary optimization technique [23]. In Stage 2, we train a deep hashing network by using the constructed hash centers as a This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23455 Figure 1. The proposed method comprises of a two-stage pipeline. Stage 1 ( left) employs a optimization procedure to produce hash centers constrained by a minimal Hamming distance dbetween any pair of hash centers, each is assigned to one image class. dis given by the Gilbert-Varshamov bound that guarantees the optimization’s feasibility. Stage 2 ( right ) employs a deep hashing network with three loss functions. The first one brings the hash code of an image close to its corresponding hash center while keeping it distant from the other centers. The second one draws similar data points within the same class even closer. The last one is to minimize quantization errors. global similarity metric. Specifically, loss functions are de- fined to make that (1) an input image’s hash code is close to its class’s hash center but is distanced from other centers, (2) the hash codes of images in the same class should be close to each other, and (3) quantization errors are minimized. The proposed method is assessed on three datasets for image retrieval. The results indicate that the obtained hash centers are always separated by the minimal distance we derived, and the proposed method outperforms the state-of- the-art deep hashing methods.
Wang_MeMaHand_Exploiting_Mesh-Mano_Interaction_for_Single_Image_Two-Hand_Reconstruction_CVPR_2023	Abstract Existing methods proposed for hand reconstruction tasks usually parameterize a generic 3D hand model or predict hand mesh positions directly. The parametric representa- tions consisting of hand shapes and rotational poses are more stable, while the non-parametric methods can predict more accurate mesh positions. In this paper, we propose to reconstruct meshes and estimate MANO parameters of two hands from a single RGB image simultaneously to utilize the merits of two kinds of hand representations. To fulfill this target, we propose novel Mesh-Mano interaction blocks (MMIBs), which take mesh vertices positions and MANO parameters as two kinds of query tokens. MMIB consists of one graph residual block to aggregate local information and two transformer encoders to model long-range depen- dencies. The transformer encoders are equipped with differ- ent asymmetric attention masks to model the intra-hand and inter-hand attention, respectively. Moreover, we introduce the mesh alignment refinement module to further enhance the mesh-image alignment. Extensive experiments on the InterHand2.6M benchmark demonstrate promising results over the state-of-the-art hand reconstruction methods.	1. Introduction Vision-based 3D hand analysis plays an important role in many applications such as virtual reality (VR) and aug- mented reality (AR). Two-hand reconstruction from a sin- gle RGB image is more challenging due to complex mutual interactions and occlusions. Besides, the skin appearance similarity makes it difficult for the network to align image features to the corresponding hand. Previous hand reconstruction works can be divided into two categories, parametric methods [3, 7, 32, 33] and non- parametric methods [4–6, 10, 18–20]. Parametric meth- ods typically learn to regress pose andshape parameters of MANO model [23], where pose represents joint rota- *Equal contribution.†Corresponding author. InputImageGTIntagHandOursFigure 1. Comparison with the state-of-the-art method IntagHand [16] for single-image two-hand reconstruction. The integration of parametric and non-parametric hand representations allows us to achieve better performance in hard cases such as severe occlusions and challenging viewpoints. tions in axis-angle representation and shape represents the coefficients of shape PCA bases. The MANO prior can yield plausible hand shapes from a single monocular image. However, they can not produce fine-grained hand meshes due to their limited capacity. With the rapid progress of graph convolutional network (GCN) and transformer techniques [10, 16, 18, 19], it is ob- served that direct mesh reconstruction can achieve state- of-the-art performance towards the Euclidean distances be- tween the ground truth vertices and the predicted vertices. Nonetheless, the non-parametric methods are less robust in handling challenging viewpoints or severe occlusions. In this paper, we introduce a novel single-image two- hand reconstruction method designed to predict mesh ver- tices positions and estimate MANO parameters simulta- neously to utilize the merits of two kinds of hand repre- sentations. The proposed Mesh-Mano interaction Hand reconstruction architecture (MeMaHand) consists of three modules: 1) the image encoder-decoder module, 2) the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 564 mesh-mano interaction module, 3) and the mesh align- ment refinement module. To extract contextually meaning- ful image features, we pre-train a classical image encoder- decoder network on auxiliary tasks including hand seg- mentation, hand 2D joints and dense mapping encodings. The low-resolution features encode more global knowledge, while the high-resolution features contain more local de- tails. Secondly, the mesh-mano interaction module stacks three mesh-mano interaction blocks (MMIBs) to transform the mesh vertices and MANO parameters queries initialized by the global image feature vector. We observe that the hand prior embedded in the MANO parameters is valuable for predicting stable hand meshes in challenging situations such as severe occlusions. MMIB consists of one graph residual block to aggregate local information and two trans- former encoders to model long-range dependencies. The transformer encoders are equipped with different asymmet- ric attention masks to model the intra-hand and inter-hand attention, respectively. Each MMIB is followed by an up- sampling operation to upsample the mesh vertices tokens in a coarse-to-fine manner. Finally, the mesh alignment re- finement module utilizes one MMIB to predict offsets for mesh vertices and MANO parameters to enhance mesh- image alignment. To improve the reliability of image ev- idence, we project mesh vertices predicted by the Mesh- Mano interaction module onto the 2D image plane. The ex- plicit mesh-aligned image features are concatenated to the transformer input tokens. The whole network, including the pre-trained image fea- ture encoder-decoder, is jointly optimized such that the im- age features better adapt to our hand mesh reconstruction task. Benefiting from the mesh-mano interaction mecha- nism and mesh alignment refinement stage, extensive ex- periments demonstrate that our method outperforms exist- ing both parametric and non-parametric methods on Inter- Hand2.6M [22] dataset. In summary, the contributions of our approach are as follows: • We propose MeMaHand to integrate the merits of para- metric and non-parametric hand representation. The mesh vertices and MANO parameters are mutually re- inforced to achieve better performance in mesh recov- ery and parameter regression. • A mesh-image alignment feedback loop is utilized to improve the reliability of image evidence. Therefore, more accurate predictions are obtained by rectifying the mesh-image misalignment. • Our method achieves superior performance on the InterHand2.6M dataset, compared with both non- parametric and parametric methods.
Wang_Learning_Conditional_Attributes_for_Compositional_Zero-Shot_Learning_CVPR_2023	Abstract Compositional Zero-Shot Learning (CZSL) aims to train models to recognize novel compositional concepts based on learned concepts such as attribute-object combinations. One of the challenges is to model attributes interacted with different objects, e.g., the attribute “wet” in “wet apple” and “wet cat” is different. As a solution, we provide anal- ysis and argue that attributes are conditioned on the recog- nized object and input image and explore learning condi- tional attribute embeddings by a proposed attribute learn- ing framework containing an attribute hyper learner and an attribute base learner. By encoding conditional at- tributes, our model enables to generate flexible attribute embeddings for generalization from seen to unseen compo- sitions. Experiments on CZSL benchmarks, including the more challenging C-GQA dataset, demonstrate better per- formances compared with other state-of-the-art approaches and validate the importance of learning conditional at- tributes. Code‡is available at https://github.com/ wqshmzh/CANet-CZSL .	1. Introduction Deep machine learning algorithms today can learn knowledge of concepts to recognize patterns. Can a ma- chine compose different learned concepts to generalize to new compositions? Compositional generalization is one of the hallmarks of human intelligence [3, 18]. To make the models equipped with this ability, Compositional Zero-Shot Learning (CZSL) [25] is proposed, where the models are trained to recognize images of unseen compositions com- posed of seen concepts. In this work, we concentrate on the situation where each composition is composed by attribute (e.g., wet) and object (e.g., apple ). For example, given im- *E-mail: wqshmzh@mail.nwpu.edu.cn †Corresponding author. E-mail: peng.wang@nwpu.edu.cn ‡Gitee: https://gitee.com/wqshmzh/canet-czsl Figure 1. The diagram of our work. We aim to learn conditional attributes conditioned on the recognized object and input image through an attribute learning framework containing an attribute hy- per learner and an attribute base learner. We first recognize the ob- ject in the input image. Then, we feed prior knowledge extracted from the conditions, which are recognized object word embedding and input image visual embedding, to the attribute hyper learner. Finally, conditional attribute embeddings are produced by the at- tribute base learner parameterized by the attribute hyper learner. ages of wet apple anddry cat , a well-trained model can rec- ognize images of new compositions dry apple andwet cat . Compositional Zero-Shot Learning of attribute-object compositions requires modeling attributes, objects, and the contextuality between them. Learning to model objects in CZSL is similar to conventional supervised object classi- fication task since the model has access to all objects in CZSL task [33]. Learning to model contextuality between attribute and object is mostly addressed in the literature [23,25,26,31,39–41]. One of the main challenges of CZSL is the appearance diversity of an attribute when composed with different objects, e.g., attribute wetinwet apple and wet cat is different. This reveals that the information of each attribute is dependent on different objects. However, most This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11197 recent works in CZSL [4, 27, 32, 33, 42, 45] extract attribute representations irrelevant to the object from seen composi- tions to infer the unseen compositions. These approaches neglect the nature of attribute diversity and learn concrete attribute representation agnostic to different objects. In this paper, we learn conditional attributes rather than learning concrete ones in a proposed Conditional Attribute Network (CANet). We first conduct analysis to determine the exact conditions by considering the recognition of at- tribute and object as computing a classification probability of attribute and object conditioned on the input image. By decomposing this probability, we demonstrate that the prob- ability of the input image belonging to an attribute is condi- tioned on the recognized object and the input image. We present an attribute learning framework to learn con- ditional attribute embeddings conditioned on the above two conditions. The framework contains an attribute hyper learner and an attribute base learner, which are sketched in Fig. 1. The attribute hyper learner learns from prior knowledge extracted from the conditions. The attribute base learner is parameterized by the attribute hyper learner and is designed to encode all attribute word embeddings into conditional attribute embeddings. With the attribute learn- ing framework, the attribute embeddings are changed along with the recognized object and input image. Finally, the at- tribute matching is processed in an attribute space where the input image embedding is projected. The attribute classifi- cation logits are computed by cosine similarities between the projected input image embedding and all conditional attribute embeddings. Additionally, we model the contex- tuality between attribute and object as composing attribute and object word embeddings. We use cosine similarities between the projected input image embedding and all com- posed attribute-object embeddings to get the classification logits. Our main contributions are as follows: • We propose to learn attributes conditioned on the rec- ognized object and input image. • We propose an attribute learning framework contain- ing an attribute hyper learner and an attribute base learner for learning conditional attribute embeddings. • Experiments and ablation studies indicate the effec- tiveness of our proposed conditional attribute network, which further validates the importance of learning con- ditional attributes in the CZSL task.
Wei_MMANet_Margin-Aware_Distillation_and_Modality-Aware_Regularization_for_Incomplete_Multimodal_Learning_CVPR_2023	Abstract Multimodal learning has shown great potentials in nu- merous scenes and attracts increasing interest recently. However, it often encounters the problem of missing modal- ity data and thus suffers severe performance degradation in practice. To this end, we propose a general framework called MMANet to assist incomplete multimodal learn- ing. It consists of three components: the deployment net- work used for inference, the teacher network transferring comprehensive multimodal information to the deployment network, and the regularization network guiding the de- ployment network to balance weak modality combinations. Specifically, we propose a novel margin-aware distilla- tion (MAD) to assist the information transfer by weigh- ing the sample contribution with the classification uncer- tainty. This encourages the deployment network to focus on the samples near decision boundaries and acquire the refined inter-class margin. Besides, we design a modality- aware regularization (MAR) algorithm to mine the weak modality combinations and guide the regularization net- work to calculate prediction loss for them. This forces the deployment network to improve its representation abil- ity for the weak modality combinations adaptively. Fi- nally, extensive experiments on multimodal classification and segmentation tasks demonstrate that our MMANet out- performs the state-of-the-art significantly. Code is available at: https://github.com/shicaiwei123/MMANet	1. Introduction Multimodal learning has achieved great success on many vision tasks such as classification [21, 33, 46], object detec- tion [26, 45, 53], and segmentation [5, 23, 41]. However, most successful methods assume that the models are trained and tested with the same modality data. In fact, limited by device [32, 39], user privacy [13, 25], and working con- dition [3, 29], it is often very costly or even infeasible toModality Customized Unified Drop RGB 10.01 11.75 -1.65 Depth 4.45 5.87 -1.42 IR 11.65 16.62 -4.97 RGB+Depth 3.41 4.61 -1.2 RGB+IR 6.32 6.68 -0.36 Depth+IR 3.54 4.95 -1.41 RGB+Depth+IR 1.23 2.21 -0.98 Table 1. The performance of customized models and the uni- fied model for different modality combinations on the CASIA- SURF dataset using the average classification error rate. The ‘customized‘ means to train a model for each combination inde- pendently while the ‘unified’ means to train only one model for all the combinations. The architectures of all the models are the same and the feature map of missing modality (such as the IR for RGB+Depth) is set as zero. collect complete modality data during the inference stage. There is thus substantial interest in assisting the incomplete or even single modality inference via the complete modality data during training. A typical solution is to reconstruct the sample or feature of the missing modalities from the available ones [10, 14, 15, 20, 29, 32]. Nevertheless, this needs to build a specific model for each modality from all possible modality combi- nations and thus has high complexity. Recent studies focus on learning a unified model, instead of a bunch of networks, for different modality combinations. Generally, many such approaches [6, 11, 12, 17, 51, 52] attempt to leverage feature fusion strategies to capture modality-invariant representa- tion so that the model can adapt to all possible modality combinations. For example, RFNet [11] designs the region- aware fusion module to fuse the features of available image modalities. Although the existing unified models are indeed able to increase the efficiency of training and deployment of the multimodal models, their performance is likely to be sub- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20039 optimal. As shown in Table 1, the customized models con- sistently outperform the unified model for different modal- ity combinations. This is because existing unified mod- els usually focus on the modality-invariant features while ignoring the modality-specific information. Note that the complementary modality-specific information of multiple modalities can help refine the inter-class discrimination and improve inference performance [2, 18, 36]. This motivates us to propose the first research question of this paper: Can a unified model consider the modality invariant and spe- cific information simultaneously while maintaining ro- bustness for incomplete modality input? To this end, we propose to guide the unified model to learn the comprehensive multimodal information from the teacher model trained with complete modality. This regular- izes the target task loss to encourage the unified model to ac- quire complementary information among different modal- ity combinations multimodal information while preserv- ing the generalization to them. Specifically, we propose a novel margin-aware distillation (MAD) that trains the uni- fied model by guiding it to mimic the inter-sample relation of the teacher model. MAD introduces the classification uncertainty of samples to re-weigh their contribution to the final loss. Since the samples near the class boundary are more likely to be misclassified and have higher classifica- tion uncertainty [8], this encourages the unified model to preserve the inter-class margin refined by the complemen- tary cues and learn the modality-specific information. Another limitation of existing unified approaches is that they struggle to obtain optimal performance for the unbal- anced training problem. To be specific, conventional multi- modal learning models tend to fit the discriminative modal- ity combination and their performance will degrade signif- icantly when facing weak modality combinations. To solve this issue, existing unified approaches introduce the auxil- iary discriminator to enhance the discrimination ability of the unimodal combinations [6, 11, 51]. This utilizes a hy- pothesis that a single modality is weaker than multiple ones. However, as shown in Table 1, no matter for the customized model or the unified model, the single Depth modality out- performs the RGB, IR, and their combinations. This indi- cates the combination with multiple weak modalities may be harder to be optimized than a single strong modality. Moreover, as shown in Table 3, RGB becomes the strong modality while Depth and IR become the weak modalities. This indicates that the modality importance is not fixed but varies with scenarios. These findings motivate us to propose the second research question: How to effectively optimize the weak modality combination in varying scenarios? To this end, we design a regularization network and MAR algorithm to assist the training of the unified network. Specifically, the regularization network generates additional predictions for all inputs. Then MAR mines and calculatesprediction loss for the sample from the weak combinations. This forces the unified model to improve its representation ability for the weak combination. In detail, MAR mines the weak combination via the memorization effect [1, 16, 49] that DNNs tend to first memorize simple examples before overfitting hard examples. As shown in Fig. 5(a), the uni- fied model tends to fit the samples containing Depth modal- ity firstly at the early stage. Therefore, MAR first mines the strong modality via the memorization effect. Then it deter- mines the combinations of rest modalities as the weak ones. Finally, we develop a model and task agnostic frame- work called MMANet to assist incomplete multimodal learning by combining the proposed MAD and MAR strate- gies. MMANet can guide the unified model to acquire com- prehensive multimodal information and balance the perfor- mance of the strong and weak modality combination si- multaneously. Extensive comparison and ablation experi- ments on multimodal classification and segmentation tasks demonstrate the effectiveness of the MMANet.
Wei_Autoregressive_Visual_Tracking_CVPR_2023	Abstract We present ARTrack , an autoregressive framework for visual object tracking. ARTrack tackles tracking as a co- ordinate sequence interpretation task that estimates object trajectories progressively, where the current estimate is in- duced by previous states and in turn affects subsequences. This time-autoregressive approach models the sequential evolution of trajectories to keep tracing the object across frames , making it superior to existing template matching based trackers that only consider the per-frame localiza- tion accuracy. ARTrack is simple and direct, eliminating customized localization heads and post-processings. Despite its simplicity, ARTrack achieves state-of-the-art performance on prevailing benchmark datasets. Source code is available athttps://github.com/MIV-XJTU/ARTrack .	1. Introduction Visual object tracking [5, 20, 34, 38, 48, 52] is a founda- tional objective in the realm of computer vision, whereby the tracker endeavors to estimate the location of an arbitrary target in each video frame, based on its initial state. Despite its ostensibly straightforward definition, the tracking task poses a significant challenge in real-world settings due to a variety of issues including but not limited to object de- formation, scale variation, occlusion, and distraction from similar objects. Fortunately, visual tracking capitalizes on abundant temporal data as its input comprises a sequence of video frames. Observationally, humans leverage temporal information to gain a perception of the target’s deformation, velocity, and acceleration trends, enabling them to maintain consistent tracking results in the face of indiscriminative or temporarily unavailable visual information. The present mainstream approaches [10,13,45,61,64] for visual object tracking typically view it as a per-frame tem- plate matching problem, neglecting the potential temporal dependencies among the video frames. These methods gen- erally follow three primary stages: (i) deep neural network- based feature extraction from the search and template images, 𝑥𝑚𝑖𝑛𝑡𝑦𝑚𝑖𝑛𝑡𝑥𝑚𝑎𝑥𝑡𝑦𝑚𝑎𝑥𝑡 Command Token Spatio -Temporal PromptsSearch Template t-2 t-1t 𝑥𝑚𝑖𝑛𝑡−2𝑦𝑚𝑖𝑛𝑡−2𝑥𝑚𝑎𝑥𝑡−2𝑦𝑚𝑎𝑥𝑡−2𝑥𝑚𝑖𝑛𝑡−1𝑦𝑚𝑖𝑛𝑡−1𝑥𝑚𝑎𝑥𝑡−1𝑦𝑚𝑎𝑥𝑡−1 …EncoderAutoregressive Decoder Figure 1. Our ARTrack framework . First, we embed visual fea- tures of template and search by an encoder. Then the coordinate tokens at current time step are interpreted by the decoder, condi- tioning on the previous estimates (as spatio-temporal prompts), and the command and visual tokens. (ii) an integration module using either convolution [2,4] or at- tention mechanisms [10,61] for feature matching/fusion, and (iii) bounding-box localization through customized heads for corner [13, 61], center/scale [64] estimation, and target classification [4, 61]. In some cases, the first two stages can be combined using a unified architecture [13, 64]. Post- processing techniques are usually employed during the local- ization step, such as Hanning window penalty [10,56,64,68] and box optimization [4, 56]. Some methods incorporate a template update mechanism to improve the target feature representation. Representative techniques in this category in- clude template image selection [61], feature integration [56], and time evolution [62, 66]. However, customized heads and post-processing techniques are complex and may require individual training and inference, which compromises the simple end-to-end framework. Moreover, tracking empha- sizes preserving localization accuracy across the sequence, while conventional per-frame training methods prioritize immediate localization accuracy, resulting in an objective mismatch between training and inference [35]. This study proposes a novel framework to visual object This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9697 tracking that differs from the mainstream methods, which typically employ a per-frame template matching task. In- stead, the authors propose to consider tracking as coordi- nate sequence interpretation , with the objective of learning a simple end-to-end model for direct trajectory estimation. The proposed approach is based on the idea that given a se- quence of frames and an initial object box, the tracker should “interpret” a sequence of coordinates that trace the object, in a manner similar to a language modeling task. The pro- posed framework models the sequential evolution of object trajectories across frames by decoding the entire trajectory sequence step by step. The current estimate is influenced by previous states and in turn influences the subsequences, thus unifying the task objectives of training and inference. Furthermore, the proposed approach simplifies the tracking pipeline by avoiding customized heads and post-processings, relying instead on direct coordinate regression. The proposed autoregressive visual tracking framework, called ARTrack, is depicted in Figure 1. The first step in this framework is to construct discrete token sequences from object trajectories using a quantization and serialization scheme [8]. The framework then adopts an encoder-decoder architecture to perceive visual information and generate tar- get sequences gradually. In this autoregressive framework, the previous outcomes serve as spatio-temporal prompts , propagating preceding motion dynamics into succeeding frames for more coherent tracking results. Notably, the model is trained using a structured loss function that maxi- mizes the likelihood of the target sequence, consistent with the task objective at test time. The authors demonstrate the ef- ficacy of this approach through extensive experiments, show- ing that the simple and neat ARTrack framework achieves state-of-the-art results on prevailing tracking benchmarks, outperforming other highly customized trackers.
Wen_DIP_Dual_Incongruity_Perceiving_Network_for_Sarcasm_Detection_CVPR_2023	Abstract Sarcasm indicates the literal meaning is contrary to the real attitude. Considering the popularity and complemen-tarity of image-text data, we investigate the task of multi-modal sarcasm detection. Di ﬀerent from other multi-modal tasks, for the sarcastic data, there exists intrinsic incon-gruity between a pair of image and text as demonstratedin psychological theories. To tackle this issue, we pro- pose a Dual Incongruity Perceiving (DIP) network con- sisting of two branches to mine the sarcastic information from factual and a ﬀective levels. F or the factual aspect, we introduce a channel-wise reweighting strategy to ob- tain semantically discriminative embeddings, and leverage gaussian distribution to model the uncertain correlationcaused by the incongruity. The distribution is generated from the latest data stored in the memory bank, which can adaptively model the di ﬀerence of semantic similarity be- tween sarcastic and non-sarcastic data. F or the a ﬀective aspect, we utilize siamese layers with shared parametersto learn cross-modal sentiment information. Furthermore,we use the polarity value to construct a relation graphfor the mini-batch, which forms the continuous contrastiveloss to acquire aﬀective embeddings. Extensive experiments demonstrate that our proposed method performs favorably against state-of-the-art approaches. Our code is released onhttps://github.com/downdric/MSD .	1. Introduction Sarcasm is an interesting and prevailing manner to ex- press users’ opinions [ 18], which means the real attitude is converse to the literal meaning [ 19]. With the develop- ment of social platforms, sarcasm detection (SD) attractsincreasing attention [ 11,40,65] due to its wide application, e.g. product review analysis, political opinion mining [ 32], etc. Automatically distinguishing sarcastic instances from ∗Equal contribution. † Corresponding author. We got a little bit ofsnow in the last Factual Incongruity (b)(a) Affective IncongruityProbabilitySuper busy at the theatre this morning Ice cream roses Foil fish packets with spinach Weather’s looking amazing todayThanks , usps ! Beautiful day with more beautiful pup A bomb made a direct hit ProbabilityX Y Z [ 0.0 0.2 0.4 0.6 0.8 1.00.20.60.8 0.4 0.0Sarcastic Non- Sarcastic XY Z[Mean: 0.65 Std: 0.93 Mean: 0.43 Std: 0.84 0.0 0.2 0.4 0.6 0.8 1.00.20.60.8 0.4 0.0XYZSarcastic Non- Sarcastic Mean: 0.53 Std: 1.27 Mean: 0.11 Std: 1.41 X Y Z [[ Figure 1. Examples from the sarcasm dataset [ 6]. (a) shows the samples (left) and statistics (right) for factual incongruity. Ac-quiring inter-modal semantic similarity 𝑆 𝑖𝑛𝑡𝑒𝑟 from CLIP [ 50], the factual incongruity is depicted by 1 −𝑆𝑖𝑛𝑡𝑒𝑟 . (b) displays the cases for aﬀective incongruity. Obtaining the models trained on FI (image) [ 70] and IMDB (text) [ 41] datasets, the incongruity is rep- resented by the diﬀerence in sentiment polarity. For both groups of samples, the top two samples are sarcastic data, and the bottomtwo samples are non-sarcastic ones. the mass of non-sarcastic content is important for any on- line service. The challenge of multi-modal sarcasm detection (MSD) mainly comes from two aspects. First, the task aims to de-tect implicit intention from data, which increases the di ﬃ- culty of learning. Speciﬁcally, compared with visual recog-nition, the expressed attitude of the sarcastic data com- monly hides in a normal stimulus and is hard to be iden- tiﬁed. Fortunately, the linguistic theory demonstrates that incongruity is an important and e ﬀective factor for sarcasm detection [ 29], which inspires researchers automatically ex- tract the positive and negative seeds [ 51]. Another challenge lies in that, while both image and text express similar infor- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2540 mation is expected in multi-modal tasks [ 3,50], this rule is not applicable to SD that discovering dissimilar informa- tion. There exists an intrinsic conﬂict between o ﬀ-the-shelf techniques for multi-modal learning and the new task in thiswork. In order to address the issue, we focus on the inter-modal incongruity for MSD. Sarcasm is a long standing topic invarious areas like psychology [ 43], sociology [ 54], and neu- robiology [ 31]. Researchers observe that sarcasm occurs when the literal meaning unexpectedly contrasts with theobserved facts [ 22,43]. The process is deﬁned as counter- factual inference [ 43]. Besides, the studies from empiri- cal theory [ 55] ﬁnd that attitude is another important factor, which is especially e ﬀective for obscure cases. In light of these theoretical works, we utilize semantic association and sentiment polarity to verify the incongruity in the sarcasmdataset [ 6]. As shown in Figure 1, the incongruity of sar- castic data is obviously larger than the non-sarcastic one infactual level, especially in terms of the mean value. Mean- while, the phenomenon also exists in the a ﬀective level. In- spired by the study and veriﬁcation above, we design ourmethod to detect the incongruity for multi-modal sarcastic data in both the factual and a ﬀective levels. We propose a Dual Incongruity Perceiving (DIP) net- work, which is consisting of Semantic Intensiﬁed Distri-bution (SID) Modeling and Siamese Sentiment Contrastive (SSC) Learning modules. In SID, based on the semantic as- sociation [ 9,44], the samples are di ﬀerentiated by an adap- tive strategy. Speciﬁcally, we maintain gaussian distribu- tions for sarcastic and non-sarcastic samples respectively, and utilize the probability generated by them to model theincongruity. Since the distributions depend on the extractedembeddings, we introduce a channel-wise reweighting strat- egy to learn representations related to sarcasm. In SSC, the aﬀective incongruity is perceived by the polarity dif- ference between the image-text pair. To e ﬃciently intro- duce sentiment information into the network, we employ two siamese layers to transmit knowledge of a ﬀective dic- tionary, i.e. SenticNet. Furthermore, with the help of the polarity intensity, the continuous contrastive learning is pro-posed to enhance the a ﬀective representations. Overall, the facutal and aﬀective information are intensiﬁed in SID and SSC, and leveraged to explicitly calculate the incongruity for MSD. Our contributions are three-fold: (1) To our knowledge, DIP is the ﬁrst work explicitly investigating and model-ing incongruity in multi-modal sarcasm detection. (2) It’s a dual perceiving network to learn sarcastic informationfrom factual and a ﬀective levels, which utilizes channel- wise reweighting and continuous contrastive strategies to acquire discriminative representations. (3) Extensive com- parisons and ablations demonstrate the e ﬀectiveness and su- periority of the proposed method.
Wang_LP-DIF_Learning_Local_Pattern-Specific_Deep_Implicit_Function_for_3D_Objects_CVPR_2023	Abstract Deep Implicit Function (DIF) has gained much popu- larity as an efﬁcient 3D shape representation. To capturegeometry details, current mainstream methods divide 3Dshapes into local regions and then learn each one with alocal latent code via a decoder . Such local methods cancapture more local details due to less diversity among local regions than global shapes. Although the diversity of localregions has been decreased compared to global approaches,the diversity in different local regions still poses a challenge in learning an implicit function when treating all regionsequally using only a single decoder . What is worse, theselocal regions often exhibit imbalanced distributions, where certain regions have signiﬁcantly fewer observations. This leads that ﬁne geometry details could not be preserved well.To solve this problem, we propose a novel Local Pattern-speciﬁc Implicit Function, named LP-DIF , to represent ashape with clusters of local regions and multiple decoders, where each decoder only focuses on one cluster of local re-gions which share a certain pattern. Speciﬁcally, we ﬁrstextract local codes for all regions, and then cluster theminto multiple groups in the latent space, where similar re-gions sharing a common pattern fall into one group. After that, we train multiple decoders for mining local patterns of different groups, which simpliﬁes the learning of ﬁne geo-metric details by reducing the diversity of local regions seen by each decoder . To further alleviate the data-imbalanceproblem, we introduce a region re-weighting module to each pattern-speciﬁc decoder using a kernel density estima-tor , which dynamically re-weights the regions during learn- *The corresponding author is Yu-Shen Liu. This work was supported by National Key R&D Program of China (2022YFC3800600), the Na-tional Natural Science Foundation of China (62272263, 62072268), andin part by Tsinghua-Kuaishou Institute of Future Media Data.ing. Our LP-DIF can restore more geometry details, and thus improve the quality of 3D reconstruction. Experiments demonstrate that our method can achieve the state-of-the- art performance over previous methods. Code is available at https://github.com/gtyxyz/lpdif.	1. Introduction Representing 3D shapes is a fundamental problem for many applications in 3D computer vision. Recently, DeepImplicit Function (DIF) [ 4,22,23,25,28,41] has gained pop- ularity for efﬁciently learning the representation of 3D ob- jects and scenes. In contrast to directly learning explicit 3D representations [ 15,30,31] (voxels, point clouds or meshes), DIF aims to train a neural network to learn the binary occu- pancy function [ 25] or signed distance function (SDF) [ 28], as given a query location and an input latent code. Such kind of representation is continuous with arbitrary precision and can handle various topology, which has achieved the state-of-the-art results in several shape reconstruction tasks. Existing DIF methods can be roughly classiﬁed into two categories: global and local approaches. Most of the earlymethods [ 4,9,20,21,25,27–29,35,37,40] fall into global approaches. These methods take advantage of one latentcode and a single decoder to represent the whole shape.Global approaches often suffer from long training time andlow reconstruction accuracy due to the limited capacity ofcapturing local geometry details. More recently, local ap-proaches [ 1,3,5,6,10–12,17,19,34,36,38] divide 3D shapes (often divide by 3D grids) into local regions and then learn each one with a local latent code via a decoder, where the decoder shares the geometric similarities among differentlocal regions. Although such local approaches can capturesome local details, a large diversity of different local regionsstill increase the difﬁculty of learning an implicit function This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21856 (d) Reference (b) Local DIF (c) Ours (a) Global DIF Figure 1. Visual comparison of 3D shape surface reconstruction. Compared with global DIF (e.g. [ 28]) and local DIF (e.g. [ 1]), our method can reconstruct the shape with ﬁne-grained geometric details. Compared with previous methods that treat all local regions equally using asingle decoder, we regard a shape as clusters of local regions and mine local patterns with different decoders. This alleviates the difﬁculty of learning caused by diverse local regions. when treating all regions equally using only a single de- coder. In addition, these local regions often exhibit imbal-anced distributions, where certain regions have signiﬁcantly fewer observations, especially in scenes. As a result, ﬁnegeometry details of shapes could not be captured well. To address the above-mentioned problems, we propose a novel Local Pattern-speciﬁc Implicit Function, named LP- DIF, for learning 3D shape representation using clusters oflocal regions with multiple decoders, where each decoderonly represents one cluster of local regions which share acertain pattern (geometric features such as facing direction,number of faces, relative positions to the region center). Speciﬁcally, we ﬁrst extract the local latent codes for alllocal regions divided by 3D grids, and then cluster them into multiple groups in the latent space, where similar re-gions sharing a common pattern fall into one group. After that, we train a separate pattern-speciﬁc decoder for each group of regions, which reduces data-imbalance among dif- ferent patterns of regions and simpliﬁes the learning of ﬁnegeometric details of 3D structures by limiting the diver- sity of regions seen to each decoder. To further alleviatethe region-imbalance problem, we introduce a region re-weighting module to each pattern-speciﬁc decoder by ker-nel density estimator, which dynamically re-weights the re-gions during learning. Our main contributions can be sum-marized as follows. • We propose a novel LP-DIF to learn local pattern- speciﬁc deep implicit function of 3D shapes for re- constructing highly detailed geometry. Compared withprevious methods that treat all local regions equally us-ing a single decoder, we regard a shape as clusters oflocal regions and mine local patterns with different de-coders. This alleviates the difﬁculty of learning caused by diverse local regions. • We introduce a dynamic region re-weighting module, which could provide more focus on less common re-gions to tackle the data-imbalance problem in each pat- tern decoder. As a result, the regions with less appear- ances can be captured more accurately. • Our method could be applied in multiple objects, sin- gle complex objects and large scale of scenes. Weimprove the state-of-the-art accuracy in surface recon- struction under various benchmarks. Figure 2illustrates the main differences between DIF, local DIF and our method. For DIF approaches, one globalcode and a decoder are used for the whole shape. For Lo-cal DIF methods, multiple local codes and a shared decoder are used. For our method, multiple clusters of regions are learned with different decoders.
Wang_3Mformer_Multi-Order_Multi-Mode_Transformer_for_Skeletal_Action_Recognition_CVPR_2023	Abstract Many skeletal action recognition models use GCNs to represent the human body by 3D body joints connected body parts. GCNs aggregate one- or few-hop graph neighbour- hoods, and ignore the dependency between not linked body joints. We propose to form hypergraph to model hyper- edges between graph nodes ( e.g., third- and fourth-order hyper-edges capture three and four nodes) which help cap- ture higher-order motion patterns of groups of body joints. We split action sequences into temporal blocks, Higher- order Transformer (HoT) produces embeddings of each temporal block based on (i) the body joints, (ii) pairwise links of body joints and (iii) higher-order hyper-edges of skeleton body joints. We combine such HoT embeddings of hyper-edges of orders 1, ..., r by a novel Multi-order Multi- mode Transformer (3Mformer) with two modules whose or- der can be exchanged to achieve coupled-mode attention on coupled-mode tokens based on ‘channel-temporal block’, ‘order- channel-body joint’, ‘channel-hyper-edge (any or- der)’ and ‘channel-only’ pairs. The ﬁrst module, called Multi-order Pooling (MP) ,additionally learns weighted ag- gregation along the hyper-edge mode, whereas the second module, Temporal block Pooling (TP) ,aggregates along the temporal block1mode. Our end-to-end trainable net- work yields state-of-the-art results compared to GCN-, transformer- and hypergraph-based counterparts.	1. Introduction Action Recognition has applications in video surveil- lance, human-computer interaction, sports analysis, and virtual reality [ 24,25,40,52–59].Di↵erent from video- based methods which mainly focus on modeling the spatio- temporal representations from RGB frames and /or opti- cal ﬂow [ 25,52–55,58], skeleton sequences, representing a spatio-temporal evolution of 3D body joints, have been *Corresponding author. 1For brevity, we write ⌧temporal blocks per sequence but ⌧varies.proven robust against sensor noises and e ↵ective in action recognition while being computationally and storage e - cient [ 24,40,52,53,56,57,59].The skeleton data is usually obtained by either localization of 2D /3D coordinates of hu- man body joints with the depth sensors or pose estimation algorithms applied to videos [ 2]. Skeleton sequences en- joy (i) simple structural connectivity of skeletal graph and (ii) temporal continuity of 3D body joints evolving in time. While temporal evolution of each body joint is highly infor- mative, embeddings of separate body joints are insensitive to relations between body parts. Moreover, while the links between adjacent 3D body joints (following the structural connectivity) are very informative as they model relations, these links represent highly correlated nodes in the sense of their temporal evolution. Thus, modeling larger groups of 3D body joints as hyper-edges can capture more complex spatio-temporal motion dynamics. The existing graph-based models mainly di ↵er by how they handle temporal information. Graph Neural Net- work (GNN) may encode spatial neighborhood of the node followed by aggregation by LSTM [ 46,65]. Alterna- tively, Graph Convolutional Network (GCN) may perform spatio-temporal convolution in the neighborhood of each node [ 64].Spatial GCNs perform convolution within one or two hop distance of each node, e.g., spatio-temporal GCN model called ST-GCN [ 64] models spatio-temporal vicin- ity of each 3D body joint. As ST-GCN applies convolution along structural connections (links between body joints), structurally distant joints, which may cover key patterns of actions, are largely ignored. ST-GCN captures ever larger neighborhoods as layers are added but su ↵ers from over- smoothing that can be mitigated by linear GCNs [ 76–78]. Human actions are associated with interaction groups of skeletal joints, e.g., wrist alone, head-wrist, head-wrist- ankles, etc. The impact of these groups of joints on each action di ↵ers, and the degree of inﬂuence of each joint should be learned. Accordingly, designing a better model for skeleton data is vital given the topology of skeleton graph is suboptimal. While GCN can be applied to a fully- connected graph ( i.e., 3D body joints as densely connected This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5620 graph nodes), Higher-order Transformer (HoT) [ 21] has been proven more e cient. Thus, we propose to use hypergraphs with hyper-edges of order 1 to rto e↵ectively represent skeleton data for ac- tion recognition. Compared to GCNs, our encoder contains an MLP followed by three HoT branches that encode ﬁrst- , second- and higher-order hyper-edges, i.e., set of body joints, edges between pairs of nodes, hyper-edges between triplets of nodes, etc. Each branch has its own learnable parameters, and processes temporal blocks2one-by-one. We notice that (i) the number of hyper-edges of Jjoints grows rapidly with order r,i.e.,⇣J i⌘ fori=1, ..., r, embed- dings of the highest order dominate lower orders in terms of volume if such embeddings are merely concatenated, and (ii) long-range temporal dependencies of feature maps are insuciently explored, as sequences are split into ⌧tempo- ral blocks for computational tractability. Merely concatenating outputs of HoT branches of orders 1 tor, and across ⌧blocks, is sub-optimal. Thus, our Multi- order Multi-mode Transformer (3Mformer) with two mod- ules whose order can be exchanged, realizes a variation of coupled-mode tokens based on ‘channel-temporal block’, ‘order- channel-body joint’, ‘channel-hyper-edge (any or- der)’ and ‘channel-only’ pairs. As HoT operates block- by-block, ‘channel-temporal block’ tokens and weighted hyper-edge aggregation in Multi-order Pooling (MP) help combine information ﬂow block-wise. Various coupled- mode tokens help improve results further due to di ↵erent focus of each attention mechanism. As the block-temporal mode needs to be aggregated (number of blocks varies across sequences), Temporal block Pooling (TP) can use rank pooling [ 13], second-order [ 14,26,33,41,60,68,80] or higher-order pooling [ 8,24,25,69,70]. In summary, our main contributions are listed as follows: i.We model the skeleton data as hypergraph of orders 1 tor(set, graph and /or hypergraph), where human body joints serve as nodes. Higher-order Transformer em- beddings of such formed hyper-edges represent various groups of 3D body joints and capture various higher- order dynamics important for action recognition. ii.As HoT embeddings represent individual hyper-edge order and block, we introduce a novel Multi-order Multi-mode Transformer (3Mformer) with two mod- ules, Multi-order Pooling and Temporal block Pool- ing. Their goal is to form coupled-mode tokens such as ‘channel-temporal block’, ‘order-channel-body joint’, ‘channel-hyper-edge (any order)’ and ‘channel-only’, and perform weighted hyper-edge aggregation and tem- poral block aggregation. 2Each temporal block enjoys a locally factored out (removed) temporal mode, which makes each block representation compact.Our 3Mformer outperforms other GCN- and hypergraph- based models on NTU-60, NTU-120, Kinetics-Skeleton and Northwestern-UCLA by a large margin.
Xu_JacobiNeRF_NeRF_Shaping_With_Mutual_Information_Gradients_CVPR_2023	Abstract We propose a method that trains a neural radiance field (NeRF) to encode not only the appearance of the scene but also semantic correlations between scene points, regions, or entities – aiming to capture their mutual co-variation pat- terns. In contrast to the traditional first-order photomet- ric reconstruction objective, our method explicitly regular- izes the learning dynamics to align the Jacobians of highly- correlated entities, which proves to maximize the mutual information between them under random scene perturba- tions. By paying attention to this second-order information, we can shape a NeRF to express semantically meaningful synergies when the network weights are changed by a delta along the gradient of a single entity, region, or even a point. To demonstrate the merit of this mutual information model- ing, we leverage the coordinated behavior of scene entities *Equal Contributions †Corresponding Author <yanchaoy@hku.hk >. The author is also af- filiated with the HKU Musketeers Foundation Institute of Data Science.that emerges from our shaping to perform label propagation for semantic and instance segmentation. Our experiments show that a JacobiNeRF is more efficient in propagating annotations among 2D pixels and 3D points compared to NeRFs without mutual information shaping, especially in extremely sparse label regimes – thus reducing annotation burden. The same machinery can further be used for entity selection or scene modifications. Our code is available at https://github.com/xxm19/jacobinerf.	1. Introduction When a real-world scene is perturbed, the response is generally local and semantically meaningful, e.g., a slight knock on a chair will result in a small displacement of just that chair. Such coherence in the perturbation of a scene evidences high mutual information between certain scene points or entities that can be leveraged to discover instances or semantic groups [37, 38]. A NeRF scene representation, however, solely supervised with 2D photometric loss may not converge to a configuration that reflects the actual scene This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16498 structure [41]; even if the density is correctly estimated, the network in general will notbe aware of the underlying se- mantic structure. As shown in Fig. 1, a perturbation on a specific entity of the scene through the network weights ac- tivates almost all other entities. This lack of semantic awareness may not be a problem for view synthesis and browsing, but it clearly is of con- cern when such neural scene representations are employed for interactive tasks that require understanding the underly- ing scene structure, e.g., entity selection, annotation prop- agation, scene editing, and so on. All these tasks can be greatly aided by a representation that better reflects the cor- relations present in the underlying reality. We take a step towards endowing neural representations with such aware- ness of the mutual scene inter-dependencies by asking how it is possible to train a NeRF, so that it not only reproduces the appearance and geometry of the scene, but also gener- ates coordinated responses between correlated entities when perturbed in the network parameter space. Current approaches that encode semantics largely treat semantic labels (e.g., instance segmentation) [17, 42] as a separate channel, in addition to density or RGB radiance. However, in the semantics case, the value of the channel (e.g., instance ID) is typically an artifact of the implemen- tation. What really matters is the decomposition of the 2D pixels (or of the scene 3D points) the NeRF encodes into groups – this is because semantics is more about rela- tionships than values. Thus, we introduce an information- theoretic technique whose goal is to “shape” an implicit NeRF representation of a scene to better reflect the underly- ing regularities (“semantics”) of the world; so as to enforce consistent variation among correlated scene pixels, points, regions, or entities, enabling efficient information propaga- tion within and across views. Thekeyto the proposed “shaping” technique is an equiv- alence between mutual information and the normalized in- ner product (cosine similarity) of the Jacobians at two pixels or 3D points. More explicitly, if we apply random delta per- turbations to the NeRF weights, the induced random values of two pixels share mutual information up to the absolute cosine similarity of their gradients or Jacobians with respect to the weights computed at the unperturbed NeRF. This the- oretical finding ensures a large correlation between scene entities with high mutual information – and thus coherent perturbation-induced behaviors – if their tangent spaces are aligned. Based on this insight, we apply contrastive learn- ing to align the NeRF gradients with general-purpose self- supervised features (e.g., DINO), which is why we term our NeRF “JacobiNeRF” . While several prior works [16, 30] distill 1st-order semantic information from 2D views to get a consensus 1st-order feature in 3D, we instead regularize the NeRF using 2nd-order, mutual information based con- trastive shaping on the NeRF gradients to achieve semanticconsensus – now encoded in the NeRF tangent space. The proposed NeRF shaping sets up resonances between correlated pixels or points and makes the propagation of all kinds of semantic information possible from sparse annota- tions – because pixels that co-vary with the annotated one are probably of the same semantics indicated by the mutual information equivalence. For example, we can use such res- onances to propagate semantic or instance information as shown in Sec. 3.4, where we also show that our contrastive shaping can be applied to gradients of 2D pixels, or of 3D points. The same machinery also enables many other func- tions, including the ability to select an entity by clicking at one of its points or the propagation of appearance edits, as illustrated in Fig. 9. Additionally, our approach suggests the possibility that a NeRF shaped with rich 2nd-order re- lational information in the way described may be capable of propagating many additional kinds of semantics without further re-shaping – because the NeRF coefficients have al- ready captured the essential “DNA” of points in the scene, of which different semantic aspects are just different expres- sions. In summary, our key contributions are: • We propose the novel problem of shaping NeRFs to re- flect mutual information correlations between scene enti- ties under random scene perturbations. • We show that the mutual information between any two scene entities is equivalent to the cosine similarity of their gradients with respect to the perturbed weights. • We develop JacobiNeRF , a shaping technique that ef- fectively encodes 2nd-order relational information into a NeRF tangent space via contrastive learning. • We demonstrate the effectiveness of JacobiNeRF with state-of-the-art performance on sparse label propagation for both semantic and instance segmentation tasks.
Woerl_Initialization_Noise_in_Image_Gradients_and_Saliency_Maps_CVPR_2023	Abstract In this paper, we examine gradients of logits of image classification CNNs by input pixel values. We observe that these fluctuate considerably with training randomness, such as the random initialization of the networks. We extend our study to gradients of intermediate layers, obtained via GradCAM, as well as popular network saliency estimators such as DeepLIFT, SHAP , LIME, Integrated Gradients, and SmoothGrad. While empirical noise levels vary, qualita- tively different attributions to image features are still pos- sible with all of these, which comes with implications for interpreting such attributions, in particular when seeking data-driven explanations of the phenomenon generating the data. Finally, we demonstrate that the observed artefacts can be removed by marginalization over the initialization distribution by simple stochastic integration.	1. Introduction Deep neural networks have revolutionized pattern recogni- tion, detecting complex structures at accuracies unheard of just a few years back. Unsurprisingly, the newly gained ability to model complex phenomena comes at costs in terms of interpretability — it is usually not obvious how nonlinear, multi-layer networks reach their conclusions. Correspondingly, a lot of research has focused on devel- oping interpretation methods for explaining how deep net- works make decisions [11], and this often takes the form ofattributing decisions to subsets of the data. In the case of image classification, this usually leads to saliency maps highlighting the image area containing decisive informa- tion [25, 26, 29, 30, 32, 35]. Strong classifiers trained from example data combined with suitable attribution methods have opened up a new ap- proach to empirical research: understanding phenomena by interpreting learned models [9]. We often know of poste- rior outcomes (for example, tumor growth rates or treata- bility with certain medication) but do not understand how these are related to prior data (say, findings from histolog- Figure 1. Logit-by-image gradients (ResNet18 on ”ImageNette” [12]). First column: reference image; second column: mean over 50 models, column 3-4: single models with random initialization. ical tissue samples). If we are able to train a strong classi- fier that can predict posterior outcomes from prior data, an attribution method could potentially explain which aspects of the data predict this outcome (for example, which visual features in the histology indicate a negative or positive ther- apeutic prognosis [38]), thereby providing new insight into the phenomenon at hand. For these kinds of research approaches, the classifier might only be an auxiliary tool: In terms of attribution, we are not primarily interested in explaining how the classifier reaches its decision (which, of course, would be highly rel- evant when studying potential data leakage or the fairness of decisions [4, 27]), but our actual goal is to accurately characterize which features in the data are related to the phenomenon to be explained. Ultimately, it is of course im- possible to be sure whether an ad-hoc classifier (even with This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1766 great statistical performance and hypothetical perfect attri- bution) actually does exploit all relevant information (and only this), but we would of course in such cases make an effort to avoid wrong or incomplete information or misattri- butions that we are already aware of. The main insight and contribution of this paper is to point out one such source of fluctuations in attributions, the im- pact of which, to the best of our knowledge, has not yet been documented in literature so far: In nonlinear CNNs, image gradients of network outputs can contain significant train- ingnoise, i.e., noise from (in particular) random weight ini- tialization and stochastic batch selection (Fig. 1). The level of such noise, i.e., information unrelated to the data itself, often exceeds the level of the attribution signal , and vari- ability includes coarse-scale variations that could suggest qualitatively varying attributions. Surprisingly, this still holds (and can even be worse) for more sophisticated at- tribution techniques, including popular approaches such as SHAP [20], LIME [25], DeepLIFT [29], or Integrated Gra- dients [35]. Even class activation maps (including top- and intermediate-level GradCAMs [26], the former to a lesser degree) can be affected by noise to an extent that could plau- sibly alter coarse-scale attribution. Exploring the phenomenon further, we observe that gra- dient noise grows with depth, is rather weak in simple con- vex architectures (linear softmax regression), and damp- ened stochastically for wide networks (as suggested by the known convexity in the infinite-width limit [7, 19]). This indicates that nonlinearity and nonconvexity might play an important role in causing the problem by either amplifying numerical noise or convergence to different local minima. We further show that training noise artifacts can be re- moved by marginalization [37], which in practice can be im- plemented with simple stochastic integration: By averaging the results of tens of independently initialized and trained networks, signal-to-noise-levels can be brought to accept- able levels. We also demonstrate that the same stochastic ensembling technique also improves the visual quality of feature visualization by optimization [23]. While marginal- ization incurs non-trivial additional computational efforts, it can remove a significant source of uncertainty when ex- plaining how data features are related to outcomes in previ- ously unknown ways.
Wang_Flow_Supervision_for_Deformable_NeRF_CVPR_2023	Abstract In this paper we present a new method for deformable NeRF that can directly use optical flow as supervision. We overcome the major challenge with respect to the compu- tationally inefficiency of enforcing the flow constraints to the backward deformation field, used by deformable NeRFs. Specifically, we show that inverting the backward deforma- tion function is actually not needed for computing scene flows between frames. This insight dramatically simplifies the problem, as one is no longer constrained to deformation functions that can be analytically inverted. Instead, thanks to the weak assumptions required by our derivation based on the inverse function theorem, our approach can be ex- tended to a broad class of commonly used backward defor- mation field. We present results on monocular novel view synthesis with rapid object motion, and demonstrate signifi- cant improvements over baselines without flow supervision.	1. Introduction Reconstructing dynamic scenes from monocular videos is a significantly more challenging task compared to its static-scene counterparts, due to lack of epipolar constraints for finding correspondences and ambiguities between mo- tion and structure. Recent advances in differentiable ren- dering have lead to various solutions using an analysis-by- synthesis strategy – solving the non-rigid deformation and structure by minimizing the difference between synthesized images and input video frames. Among those, deformable neural radiance fields [14, 21, 25, 31] has been a notable technique to represent dynamic scenes and shows plausi- ble space-time view synthesis results. However, the current implementations only warrant success on teleporting-like videos whose camera motions are significantly more rapid than object motions. Quality of their results significantly decrease on videos with more rapid object motions [6]. In this work, we conjecture the deficiency of these de- formable NeRF-based methods is mainly due to lack of withflow supervisionw/oflow supervision inputsFigure 1. We propose a method to use optical flow supervision for deformable NeRF. It noticeably improves novel view synthesis for monocular videos with rapid object motions. In the figure, we visualize rendered novel view images and depth maps for the first and last frame of the input video. temporal regularization. As they represent deformation as backward warping from the sampled frames to some canon- ical space, the motions or scene flows between temporally adjacent frames is not directly modeled nor supervised. An- other deficiency is that these methods minimize photomet- ric error alone, which is insufficient for gradient descent to overcome poor local minima when the canonical and other frames has little spatial overlap. This deficiency is severe for non-object-centric videos with large translations. The community has explored optical flow as an addi- tional cue to help supervise the temporal transitions of other motion representations, such as scene flow fields [4, 5, 13, 43] and blend skinning fields [40]. However, enforcing flow constraints with respect to a generic backward warp- ing field as in Nerfies [21] is non-trivial. Intuitively, to com- pute scene flows, it requires inverting the backward warp by having a forward warp which maps points from canonical space to other frames. Then scene flows can be computed either by taking time derivative of the forward warp or pre- dicting the next position of a point through evaluating the forward warp. But this can be problematic since analytical inverse of a complicated non-bijective function ( e.g. neu- ral networks) is impossible, and an approximate solution by This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21128 having an auxiliary network to represent the forward warp will introduce computational overhead and is theoretically ill-posed. Counter to this intuition, we will show that invert- ing the backward warping function is actually not needed for computing scene flows between frames. The main contribution of this paper is: we derive an analytical solution to compute velocities of objects directly from the backward warping field of the deformable NeRF. The velocities are then used to compute scene flows through temporal integration, which allows us to supervise the de- formable NeRF through optical flow. This leads to signif- icant improvement on videos with more rapid motions, as shown in Fig. 1. The advantage of our approach is twofold: (i) Our method applies to all kinds of backward warping function, thanks to the weak assumptions required by the inverse function theorem which our derivation is based on. Thus our method is more general compared to other works using invertible normalizing flows [11] or blend skinning [3, 40]. (ii) Our method is also computationally more tractable com- pared to neural scene flow fields [4, 12, 13], which would require integrating flows over long period of time to reach some canonical frame.
Wu_Sparsely_Annotated_Semantic_Segmentation_With_Adaptive_Gaussian_Mixtures_CVPR_2023	Abstract Sparsely annotated semantic segmentation (SASS) aims to learn a segmentation model by images with sparse labels (i.e., points or scribbles). Existing methods mainly focus on introducing low-level affinity or generating pseudo la- bels to strengthen supervision, while largely ignoring the inherent relation between labeled and unlabeled pixels. In this paper, we observe that pixels that are close to each other in the feature space are more likely to share the same class. Inspired by this, we propose a novel SASS frame- work, which is equipped with an Adaptive Gaussian Mix- ture Model (AGMM). Our AGMM can effectively endow re- liable supervision for unlabeled pixels based on the distri- butions of labeled and unlabeled pixels. Specifically, we first build Gaussian mixtures using labeled pixels and their relatively similar unlabeled pixels, where the labeled pix- els act as centroids, for modeling the feature distribution of each class. Then, we leverage the reliable information from labeled pixels and adaptively generated GMM predic- tions to supervise the training of unlabeled pixels, achieving online, dynamic, and robust self-supervision. In addition, by capturing category-wise Gaussian mixtures, AGMM en- courages the model to learn discriminative class decision boundaries in an end-to-end contrastive learning manner. Experimental results conducted on the PASCAL VOC 2012 and Cityscapes datasets demonstrate that our AGMM can establish new state-of-the-art SASS performance. Code is available at https://github.com/Luffy03/AGMM-SASS.	1. Introduction Semantic segmentation [2, 8, 42] aims to assign the corresponding pixel-wise semantic labels for a given im- age, which is a fundamental computer vision task. Pre- *Corresponding author Figure 1. (a) Illustration of SASS task. (b) Different from existing SASS frameworks, our AGMM leverages the reliable information of labeled pixels and generates GMM predictions for dynamic on- line supervision. fdenotes the model, PandGrepresent segmen- tation and GMM predictions, respectively. Solid and dashed lines represent model propagation and supervision, respectively. vious deep learning based semantic segmentation meth- ods [3, 9, 43] trained on large amounts of data with accu- rate pixel-wise annotations have demonstrated outstanding achievements. However, collecting such dense annotations always requires cumbersome manual efforts, which heav- ily limits the development of semantic segmentation meth- ods. To reduce the cost of manual annotations, many re- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15454 Figure 2. (a) Observation of the inherent relation between the labeled and unlabeled pixels. (b) Category-wise performance on the PASCAL VOC 2012 dataset. The black line, blue bar, and orange bar represent the IoU of all unlabeled pixels, unlabeled pixels that are similar to labeled pixels, and unlabeled pixels that are dissimilar to labeled pixels, respectively. σis the variance of a class (Eq. 5). cent works [4,17,27,29,47,48,59] have been made towards sparsely annotated semantic segmentation (SASS), which learns segmentation models via sparse labels, i.e., points or scribbles, as shown in Fig. 1(a). The sparse annotations are cheap to obtain and also contain the least necessary category and location information. Thus, SASS has high research potential in terms of the trade-off between information and costs. The main challenge of SASS is the lack of information for supervision. Existing SASS methods can be roughly divided into three categories, i.e., low-level regulariza- tion [26,30,34,47,48], pseudo supervision [7,27,35,59,60], and consistency learning [19, 38], as shown in Fig. 1(b). Specifically, the low-level regularization methods [26, 30, 34, 47, 48] focus on introducing the low-level affinity of the raw images for supervision. However, the low-level infor- mation is not reliable enough to be associated with the high- level semantics. Pseudo supervision [27, 35, 59, 60] aims to generate pseudo labels via training with sparse labels, and then uses these pseudo labels to learn a more robust segmentation model. However, it commonly requires time- consuming multi-stage training and the generated pseudo labels are always coarse and ambiguous, which significantly hinders the learning of unlabeled pixels. Consistency learn- ing [19, 38, 55] further proposes to learn consistent repre- sentations in the high-dimension feature space, but it cannot directly supervise the final predictions at the category level. To solve these problems, we aim to address the SASS task with more reliable supervision. To this end, we argue that the reliable information of labeled pixels should be fur- ther exploited. Previous methods only employ the labeled pixels for partial cross-entropy supervision, while largely ignoring the inherent relation between labeled and unla-beled pixels. As illustrated in Fig. 2, we observe that the similarity between labeled and unlabeled pixels is highly as- sociated with the predictions of unlabeled pixels. As shown in Fig. 2(a), if an unlabeled pixel is similar to the labeled pixel in the feature space, its corresponding prediction is more likely to be consistent with the category of the labeled pixel. In Fig. 2(b), we calculate the distance d(see Eq. 6) between labeled and unlabeled pixels to measure the simi- larity, i.e.,d < σ as similar and d > σ as not similar. It can be seen that the similarity between labeled and unla- beled pixels is highly associated with the accuracy of the predictions. To this end, we propose to explicitly leverage the similarity between the labeled and unlabeled pixels to generate supervision information. The key challenge is how to effectively model the similarity between the labeled and unlabeled pixels. In this paper, we propose a novel Adaptive Gaussian Mixture Model (AGMM) framework, which is realized by incorporating a GMM branch into the traditional segmenta- tion branch. Specifically, we assign the labeled pixels as the centroids of Gaussian mixtures, enabling us to model the data distribution of each class in the high-dimension fea- ture space. Each Gaussian mixture represents the distribu- tion of a class, which consists of the centered labeled pix- els and the relatively similar unlabeled pixels. In this way, we build a GMM to measure the feature similarity between labeled and unlabeled pixels, producing soft GMM predic- tions to supervise the unlabeled regions from a probabilis- tic perspective. The process of GMM formulation works in an adaptive manner, where the parameters of GMM are dynamically adapted to the input features, achieving end-to- end online self-supervision. The GMM branch is progres- sively optimized during training, enabling us to learn more 15455 discriminative Gaussian mixtures adaptively. There are three appealing advantages in our proposed AGMM. First, by capturing category-wise Gaussian mix- tures for feature representations, we can learn discrimina- tive decision boundaries between different classes via very limited supervision. Second, AGMM pushes each unla- beled pixel into or away from specific category-wise Gaus- sian mixtures, which further enables an end-to-end con- trastive representation learning. Finally, we leverage the reliable information from labeled pixels to generate GMM predictions for the unlabeled pixels, achieving more reliable supervision. We conduct experiments under the point- and scribble- supervised settings on two widely used datasets, i.e., PAS- CAL VOC 2012 [14] and Cityscapes [12]. It is worth noting that compared with existing SASS methods, our AGMM does not require extra information for supervision [19,26,30,34,50], multi-stage training [7,35,37,59,60], and time-consuming post-processing [27, 31, 50, 60]. Extensive experiments demonstrate that our AGMM outperforms the existing state-of-the-art SASS methods.
Wu_EDA_Explicit_Text-Decoupling_and_Dense_Alignment_for_3D_Visual_Grounding_CVPR_2023	Abstract 3D visual grounding aims to find the object within point clouds mentioned by free-form natural language descrip- tions with rich semantic cues. However, existing methods either extract the sentence-level features coupling all words or focus more on object names, which would lose the word- level information or neglect other attributes. To alleviate these issues, we present EDA that Explicitly Decouples the textual attributes in a sentence and conducts Dense Alignment between such fine-grained language and point cloud objects. Specifically, we first propose a text decou- pling module to produce textual features for every seman- tic component. Then, we design two losses to supervise ∗Corresponding author . This work was supported in part by Shenzhen Research Project under Grant JCYJ20220531093215035.the dense matching between two modalities: position align- ment loss and semantic alignment loss. On top of that, we further introduce a new visual grounding task, locat- ing objects without object names, which can thoroughly evaluate the model’s dense alignment capacity. Through experiments, we achieve state-of-the-art performance on two widely-adopted 3D visual grounding datasets, Scan- Refer and SR3D/NR3D, and obtain absolute leadership on our newly-proposed task. The source code is available at https://github.com/yanmin-wu/EDA .	1. Introduction Multi-modal cues can highly benefit the 3D environ- ment perception of an agent, including 2D images, 3D point clouds, and language. Recently, 3D visual grounding (3D VG) [8,50], also known as 3D object referencing [3], has at- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19231 tached much attention as an important 3D cross-modal task. Its objective is to find the target object in point cloud scenes by analyzing the descriptive query language, which requires understanding both 3D visual and linguistic context. Language utterances typically involve words describ- ing appearance attributes, object categories, spatial rela- tionships and other characteristics, as shown by different colours in Fig. 1(a), requiring that the model integrate multi- ple cues to locate the mentioned object. Compared with 2D Visual Grounding [18,19,67,71], the sparseness and incom- pleteness of point clouds, and the diversity of language de- scriptions produced by 3D multi-view, make 3D VG more challenging. Existing works made significant progress from the following perspectives: improving point cloud features extraction by sparse convolution [70] or 2D images assis- tance [68]; generating more discriminative object candi- dates through instance segmentation [33] or language mod- ulation [46]; identifying complex spatial relationships be- tween entities via graph convolution [23] or attention [6]. However, we observe two issues that remain unex- plored. 1) Imbalance : The object name can exclude most candidates, and even in some cases, there is only one name-matched object, as the “door” and“refrigerator” in Fig. 1(b1, b2). This shortcut may lead to an inductive bias in the model that pays more attention to object names while weakening other properties such as appearance and relation- ships, resulting in imbalanced learning. 2) Ambiguity : Ut- terances frequently refer to multiple objects and attributes (such as “black object, tall shelf, fan” in Fig. 1(b4)), while the model’s objective is to identify only the main object, leading to an ambiguous understanding of language de- scriptions. These insufficiencies of existing works stem from their characteristic of feature coupling and fusing im- plicitly. They input a sentence with different attribute words but output only one globally coupled sentence-level fea- ture that subsequently matches the visual features of can- didate objects. The coupled feature is ambiguous because some words may not describe the main object (green text in Fig. 1) but other auxiliary objects (red text in Fig. 1). Alternatively, using the cross-modal attention of the Trans- former [21, 63] automatically and implicitly to fuse visual and text features. However, this may encourage the model to take shortcuts, such as focusing on object categories and ignoring other attributes, as previously discussed. Instead, we propose a more intuitive decoupled and ex- plicit strategy. First, we parse the input text to decouple different semantic components, including the main object word, pronoun, attributes, relations, and auxiliary object words. Then, performing dense alignment between point cloud objects and multiple related decoupled components achieves fine-grained feature matching, which avoids the inductive bias resulting from imbalanced learning of differ- ent textual components. As the final grounding result, weexplicitly select the object with the highest similarity to the decoupled text components (instead of the entire sentence), avoiding ambiguity caused by irrelevant components. Ad- ditionally, to explore the limits of VG and examine the com- prehensiveness and fine-graininess of visual-language per- ception of the model, we suggest a challenging new task: Grounding without object name (VG-w/o-ON) , where the name is replaced by “object” (see Fig. 1(b)), forcing the model to locate objects based on other attributes and relationships. This setting makes sense because utterances that do not mention object names are common expressions in daily life, and in addition to testing whether the model takes shortcuts. Benefiting from our text decoupling oper- ation and the supervision of dense aligned losses, all text components are aligned with visual features, making it pos- sible to locate objects independent of object names. To sum up, the main contributions of this paper are as follows: 1)We propose a text decoupling module to parse linguistic descriptions into multiple semantic components, followed by suggesting two well-designed dense aligned losses for supervising fine-grained visual-language feature fusion and preventing imbalance and ambiguity learning. 2)The challenging new 3D VG task of grounding without object names is proposed to comprehensively examine the model’s robust performance. 3)We achieve state-of-the-art performance on two datasets (ScanRefer and SR3D/NR3D) on the regular 3D VG task and absolute leadership on the new task evaluated by the same model without retraining.
Xu_Learning_Open-Vocabulary_Semantic_Segmentation_Models_From_Natural_Language_Supervision_CVPR_2023	Abstract This paper considers the problem of open-vocabulary se- mantic segmentation (OVS), that aims to segment objects of arbitrary classes beyond a pre-defined, closed-set cat- egories. The main contributions are as follows: First , we propose a transformer-based model for OVS, termed as OVSegmentor, which only exploits web-crawled image- text pairs for pre-training without using any mask annota- tions. OVSegmentor assembles the image pixels into a set of learnable group tokens via a slot-attention based bind- ing module, then aligns the group tokens to correspond- ing caption embeddings. Second , we propose two proxy tasks for training, namely masked entity completion and cross-image mask consistency. The former aims to infer all masked entities in the caption given group tokens, that enables the model to learn fine-grained alignment between visual groups and text entities. The latter enforces consis- tent mask predictions between images that contain shared entities, encouraging the model to learn visual invariance. Third , we construct CC4M dataset for pre-training by fil- tering CC12M with frequently appeared entities, which sig- nificantly improves training efficiency. Fourth , we per- form zero-shot transfer on four benchmark datasets, PAS- CAL VOC, PASCAL Context, COCO Object, and ADE20K. OVSegmentor achieves superior results over state-of-the- art approaches on PASCAL VOC using only 3% data (4M vs 134M) for pre-training.	1. Introduction Semantic segmentation considers the problem of assign- ing class labels to each pixel in the image. It plays criti- cal roles in a wide range of real-world scenarios, including autonomous driving, computer-aided diagnosis and satellite image analysis, to name a few. Generally speaking, two lines of research dominate semantic segmentation, one way is to cluster the pixels into different groups and assign a *Corresponding author Seg ModelClosed-set semantic segmentation Open-vocabulary semantic segmentation (ours){cat, dog} {cat, dog, fire hydrant } Seg Model Training Inference A train arrives at its destination.Three cats are sleeping on the bed.Visual encoder Text encoder A photo of {cat/dog/fire hydrant} Visual encoder Text encoder Figure 1. An illustration of open-vocabulary semantic segmen- tation. Models trained on closed-set classes (cat and dog) fail to segment novel class (fire hydrant). We train a visual encoder and a text encoder on web-collected image-text pairs without using any mask labels, and our model can segment arbitrary object classes. semantic label to each group; the other idea treats segmen- tation as pixel-wise classification, casting each pixel into one category. Despite tremendous progress, the scalability of existing approaches that rely on supervised training has been fundamentally limited: (1) costly annotation proce- dure. Extensive manual pixel-wise annotations are required for training segmentation models; (2) closed-set segmenta- tion. The model is restricted to segmenting objects from a closed-set of categories. Whenever a new dataset comes, the model requires re-training. In this paper, our goal is to train an open-vocabulary semantic segmentation (OVS) model, by exploiting freely available image-caption pairs on Internet, as illustrated in Fig. 1. The recent work, for example, CLIP [41] and ALIGN [23] have demonstrated that a combination of large-scale image-caption pairs and simple noise con- trastive estimation can learn powerful image-text embed- dings from scratch, and show strong “zero-shot” general- ization abilities for open-vocabulary classification. Addi- tionally, GroupViT [53] extends the idea towards semantic segmentation by training a segmentation model with text supervision only. They perform hierarchical grouping of This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2935 visual tokens, which are then aligned to the correspond- ing text embeddings via a contrastive loss. However, the following issues remain challenging and unsolved: First , the captions only provide coarse, image-level descriptions, which are insufficient for training semantic segmentation models where fine-grained, pixel-wise supervision is usu- ally needed. Second , the diversity of web-collected data is large, that requires the model to learn visual invariance on objects of interest, with only weak supervision provided. For instance, the visual appearance of two images with sim- ilar captions can be drastically different. To tackle the above challenges, (i) we propose a transformer-based model for open-vocabulary semantic segmentation, dubbed as OVSegmentor, that can segment objects of arbitrary categories via zero-shot transfer, with only image-caption pairs for pre-training. Specifically, we introduce learnable group tokens to cluster image patches via a slot-attention [33] based binding module, and align the group tokens with corresponding caption embeddings. Note that our model neither requires ground-truth masks for training nor additional re-training on target segmenta- tion datasets, substantially alleviating the annotation efforts and improving transfer efficiency; (ii) As for training on the image-caption dataset, we propose two proxy tasks, namely masked entity completion and cross-image mask consis- tency, the former trains the model to infer allthe masked entities in the sentence given the group tokens, and the lat- ter enforces consistent mask prediction for images with the common entity. Both tasks have shown to be beneficial in learning entity-specific, fine-grained and visually invari- ant group semantics; (iii) We construct an image-caption dataset, termed as CC4M, by designing an automatic ap- proach to filter CC12M [7] with frequently appeared visual entities, significantly improving the training efficiency. We pre-train the proposed OVSegmentor on our filtered image-caption dataset (CC4M), without using any man- ual segmentation masks whatsoever. The model is eval- uated on four segmentation benchmarks, PASCAL VOC 2012 [18], PASCAL Context [38], COCO Object [31], and ADE20K [59] in a zero-shot manner, i.e., the model is directly evaluated on target datasets without any finetun- ing. Extensive experiments demonstrate that our model sur- passes existing models that are trained with full supervision and outperforms state-of-the-art self-supervised approaches on PASCAL VOC by using only 3% data (4M vs 134M) for pre-training, significantly improving the training efficiency.
Wang_Image_as_a_Foreign_Language_BEiT_Pretraining_for_Vision_and_CVPR_2023	Abstract A big convergence of language, vision, and multimodal pretraining is emerging. In this work, we introduce a general-purpose multimodal foundation model BEIT-3, which achieves excellent transfer performance on both vi- sion and vision-language tasks. Specifically, we advance the big convergence from three aspects: backbone architec- ture, pretraining task, and model scaling up. We use Mul- tiway Transformers for general-purpose modeling, where the modular architecture enables both deep fusion and modality-specific encoding. Based on the shared back- bone, we perform masked “language” modeling on images (Imglish ), texts (English), and image-text pairs (“parallel sentences”) in a unified manner. Experimental results show thatBEIT-3obtains remarkable performance on object de- tection (COCO), semantic segmentation (ADE20K), image classification (ImageNet), visual reasoning (NLVR2), visual question answering (VQAv2), image captioning (COCO), and cross-modal retrieval (Flickr30K, COCO).	1. Introduction: The Big Convergence Recent years have featured a trend toward the big con- vergence of language [14, 15, 46], vision [3, 43], and mul- timodal [45, 62, 69] pretraining. By performing large-scale pretraining on massive data, we can easily transfer the mod- els to various downstream tasks. It is appealing that we can pretrain a general-purpose foundation model that handles multiple modalities. In this work, we advance the conver- gence trend for vision-language pretraining from the fol- lowing three aspects. First, the success of Transformers [59] is translated from language to vision [16] and multimodal [26, 62] problems. The unification of network architectures enables us to seam- lessly handle multiple modalities. For vision-language modeling, there are various ways to apply Transformers *Equal contribution. †Corresponding author. V-FFN Vision ExpertL-FFN VL-FFN 𝐿x Multimodal InputShared Multi -Head Self-AttentionSwitching Modality Experts Language ExpertVL Expert BEiT -3 (Multiway Transformer ) Images TextsImage -Text PairsMasked Data ModelingFigure 1. Overview of BE IT-3 pretraining. We perform masked data modeling on monomodal (i.e., images, and texts) and multi- modal (i.e., image-text pairs) data with a shared Multiway Trans- former as the backbone network. due to the different natures of downstream tasks. For ex- ample, the dual-encoder architecture is used for efficient retrieval [45], encoder-decoder networks for generation tasks [63], and the fusion-encoder architecture for image- text encoding [26]. However, most foundation models have to manually convert the end-task formats according to the specific architectures. Moreover, the parameters are usu- ally not effectively shared across modalities. In this work, we adopt Multiway Transformers [62] for general-purpose modeling, i.e., one unified architecture shared for various downstream tasks. The modular network also compre- hensively considers modality-specific encoding and cross- modality fusion. Second, the pretraining task based on masked data mod- eling has been successfully applied to various modalities, such as texts [14] and images [3, 43]. Current vision- language foundation models usually multitask other pre- training objectives (such as image-text matching), render- ing scaling-up unfriendly and inefficient. In contrast, we only use one pretraining task, i.e., mask-then-predict, to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19175 train a general-purpose multimodal foundation model. By regarding the image as a foreign language (i.e., Imglish ), we handle texts and images in the same manner without funda- mental modeling differences. Consequentially, image-text pairs are utilized as “parallel sentences” in order to learn the alignments between modalities. We also show that t
Wang_MDL-NAS_A_Joint_Multi-Domain_Learning_Framework_for_Vision_Transformer_CVPR_2023	Abstract In this work, we introduce MDL-NAS, a unified frame- work that integrates multiple vision tasks into a manage- able supernet and optimizes these tasks collectively un- der diverse dataset domains. MDL-NAS is storage-efficient since multiple models with a majority of shared parame- ters can be deposited into a single one. Technically, MDL- NAS constructs a coarse-to-fine search space, where the coarse search space offers various optimal architectures for different tasks while the fine search space provides fine- grained parameter sharing to tackle the inherent obstacles of multi-domain learning. In the fine search space, we sug- gest two parameter sharing policies, i.e., sequential shar- ing policy and mask sharing policy. Compared with pre- vious works, such two sharing policies allow for the par- tial sharing and non-sharing of parameters at each layer of the network, hence attaining real fine-grained parameter sharing. Finally, we present a joint-subnet search algorithm that finds the optimal architecture and sharing parameters for each task within total resource constraints, challeng- ing the traditional practice that downstream vision tasks are typically equipped with backbone networks designed for image classification. Experimentally, we demonstrate that MDL-NAS families fitted with non-hierarchical or hierar- chical transformers deliver competitive performance for all tasks compared with state-of-the-art methods while main- taining efficient storage deployment and computation. We also demonstrate that MDL-NAS allows incremental learn- ing and evades catastrophic forgetting when generalizing to a new task.	1. Introduction Recently, transformers have become the standard pattern for natural language processing (NLP) tasks due to their efficacy in modelling long-range relationships via the self- *Corresponding author. Output1 Layer L Layer 3 Layer 2 Layer 1 Many -to-oneLayer L Layer 3 Layer 2 Layer 1 Many -to-many(MDL)Layer L Layer 3 Layer 2 Layer 1 One-to-manyIutput1Multiple dataset domainsMultiple tasksFigure 1. Illustration of differences between multi-domain learn- ing (MDL) and other learning paradigms. MDL-NAS jointly op- timizes multiple vision tasks under different dataset domains. L denotes the layer numbers in the backbone. attention mechanism [41]. Such success and good prop- erties of transformers have spawned a slew of subsequent works that apply them to a wide variety of computer vision tasks, such as image classification [6,27,32,49], object de- tection [5, 57], semantic segmentation [55], and video un- derstanding [56], achieving impressive results. However, these methods only apply transformer to a specific domain. After observing the success of transformers, a naive ques- tion arises: could a transformer simultaneously handle mul- tiple vision tasks under a variety of dataset domains ? While a few works have investigated the usage of trans- formers to handle multiple input modalities (i.e., images and text), they typically concentrate on a particular task, such as visual question answering [20, 24], i.e., many-to- one mapping. Also, some methods [2,26] explore to simul- taneously performing depth estimation, surface normal esti- mation, and semantic segmentation on a given input image. However, these methods are restricted to a single-domain setting, where all inputs are the same, i,e,, one-to-many mapping. In addition, there are some works [19, 28] that employ transformers to solve different tasks under multiple domains (multi-domain learning), which is more realistic, i.e., many-to-many mapping, as shown in Fig. 1. Never- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20094 All-share Spe-norm Spe-attn Spe-ffn All-spe60657075808590Top1 Acc.(%)80.8 80.8 81.0 81.0 81.381.9 82.0 82.1 82.2 82.5Swin transformer Autoformer(a) Image classification All-share Spe-norm Spe-attn Spe-ffn All-spe0102030405060mAP(%)39.0 39.441.242.443.244.1 44.3 44.8 45.2 44.9Swin transformer Autoformer (b) Object detection All-share Spe-norm Spe-attn Spe-ffn All-spe010203040506070mIoU(%)45.8 45.5 44.7 44.5 43.948.1 48.0 47.3 46.7 47.1Swin transformer Autoformer (c) Semantic segmentation Figure 2. Adjust baselines that shares all parameters in backbone with task-specific normalization (Spe-norm), task-specific self-attention (Spe-attn), task-specific feed forward network (Spe-ffn) and all task-specific parameters (All-spe) under the same training recipe. theless, these methods utilize diverse encoders to manage different dataset domains, which is inefficient in terms of storage deployment. In this work, we investigate a unified network that optimizes multiple vision tasks over multiple dataset domains to enable all tasks to share as many param- eters as feasible while maintaining promising performance. As a preliminary step, we conduct experiments to observe the performance impact of treating various components of vision transformers as task-specific parameters. As depicted in Fig. 2, considering the multihead self- attention (MHSA) layer or feed-forward network (FFN) or LayerNorm (LN) throughout the backbone as task-specific parameters can all achieve a certain performance gain for classification and detection tasks over a baseline that shares all parameters. Besides, we observe that the performance of semantic segmentation is elevated when all parameters are shared, indicating that closely-related tasks have mu- tual benefits whereas some tasks have conflicts against each other under multi-domain learning setting. Consequently, to use task-shared parameters for learning task-reciprocal features while using task-specific parameters for mitigating conflicts, sharing parameters with various proportions in- side each layer is an immediate thought, which motivates us to find a method to supply different share ratios for different layers in the network. Moreover, when optimizing multiple tasks collectively, we typically equip these tasks with the backbone designed for image classification, which may be sub-optimal due to the gap between the image classification task and other vision tasks. To tackle these issues, we introduce MDL-NAS, a uni- fied framework based on vision transformers, which accom- modates multiple vision tasks under heterogeneous dataset domains into a modest supernet and jointly optimizes these tasks. Specifically, we first construct a coarse search space comprising embedding dimension, heads number, query/key/value dimension, and MLP ratios for each trans- former block to discover different optimal architectures for diverse tasks. Moreover, such space comprises candidate ar- chitectures with a wide spectrum of model size, which pro- vides certain flexibility for final model deployment. Basedon the coarse search space, we design a fine search space that offers fine-grained parameter sharing for all tasks to resolve the inherent challenges of multi-domain learning. In the fine search space, we suggest two parameter sharing policies, namely sequential sharing policy and mask shar- ing policy. Sequential sharing policy enables all tasks to share parameters for each layer in order, which allows to customize the parameter share ratio. Mask sharing policy provides maximum flexibility for different tasks to share pa- rameters with various proportions and channels inside each layer. Following Autoformer [6], to address the efficiency issue, we leverage the weight entanglement training strat- egy to train MDL-NAS, allowing thousands of subnets to be extremely well-trained. During the search stage, we propose a joint-subnet search algorithm that finds the optimal architecture and sharing parameters for each task under total resource con- straints. The searched subnets with various architectures share as many parameters as possible in the backbone, guar- anteeing excellent performance for each task while keeping storage-efficient for model deployment. Experiments show that the searched models with weights inherited from the supernet outperform several baselines and are comparable with the state-of-the-art methods that are trained individually for specific tasks. We also demon- strate that MDL-NAS allows incremental learning and evades catastrophic forgetting when generalizing to a new task. Thus, MDL-NAS is more parameter-efficient and can scale up more gracefully with the number of tasks increas- ing, as illustrated in Sec. 4.4. The key contributions of this work can be summarized as: (1) We propose MDL-NAS that accepts multiple dataset domains as input to optimize multiple vision tasks concur- rently. (2) We construct a coarse-to-fine search space, with the coarse search space finding optimal architectures for all tasks and the fine search space coupled with sequential or mask sharing policy providing fine-grained shared parame- ters to learn task-reciprocal features and extra task-specific parameters for learning task-related features. (3) We intro- duce a subnet search algorithm to jointly search architec- 20095 tures and share ratios, enabling all tasks to share as many parameters as feasible while ensuring high performance for each task. (4) We demonstrate that MDL-NAS allows in- cremental learning with fewer parameters.
Wang_FEND_A_Future_Enhanced_Distribution-Aware_Contrastive_Learning_Framework_for_Long-Tail_CVPR_2023	Abstract Predicting the future trajectories of the traffic agents is a gordian technique in autonomous driving. However, tra- jectory prediction suffers from data imbalance in the preva- lent datasets, and the tailed data is often more complicated and safety-critical. In this paper, we focus on dealing with the long-tail phenomenon in trajectory prediction. Previ- ous methods dealing with long-tail data did not take into account the variety of motion patterns in the tailed data. In this paper, we put forward a future enhanced contrastive learning framework to recognize tail trajectory patterns and form a feature space with separate pattern clusters. Fur- thermore, a distribution aware hyper predictor is brought up to better utilize the shaped feature space. Our method is a model-agnostic framework and can be plugged into many well-known baselines. Experimental results show that our framework outperforms the state-of-the-art long-tail pre- diction method on tailed samples by 9.5% on ADE and 8.5% on FDE, while maintaining or slightly improving the aver- aged performance. Our method also surpasses many long- tail techniques on trajectory prediction task.	1. Introduction Trajectory prediction is of great importance in au- tonomous driving scenarios [27]. It aims to predict a series of future positions for the agents on the road given the ob- served past tracks. There have been many recent methods in trajectory prediction, both unimodal [1, 48] and multi- modal [10, 37, 38, 49]. Despite the high accuracy those prediction methods have achieved, most of them treat the samples in the datasets equally in both training and evaluation phases. But there is a long-tailed phenomenon in prevalent datasets [28]. For *Equal contributions. †Corresponding author. Figure 1. The long-tailed final displacement errors of the state- of-the-art prediction network: Trajectron++ EWTA [28] on ETH- UCY . The long-tail part of the dataset contains various compli- cated motion patterns, and predicting them is challenging. example, in real traffic scenes, most of the trajectories fol- low certain simple kinematic rules, while deviating and collision-avoiding circumstances are scarce. Therefore, the frequent cases are often simple and easy to predict, while the tail cases are often complicated with many motion pat- terns and suffer from large prediction errors, which makes them more safety-critical, as shown in Fig. 1 for the univ dataset. Despite of its significance, the long-tail prediction problem have been rarely discussed in literature. It has been pointed out that the feature encoders largely suffer from long-tail data. In the training process, the head samples are encountered more often and dominate the la- tent space, while the tailed samples will be modeled insuf- ficiently, as discussed in [24, 28, 39]. Feature embeddings of the tailed data can even be mixed up with the ones of the head data as discussed in [28], therefore the performances of the tailed samples could be harmed. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1400 In this paper, we pick up the general idea of using con- trastive learning to enhance the model ability on long-tailed data. A new framework is developed called FEND: Future ENhanced Distribution-aware contrastive trajectory predic- tion, which is a pattern-based contrastive feature learning framework enhanced by future trajectory information. An offline trajectory clustering process and prototypical con- trastive learning are introduced for recognizing and sepa- rating different trajectory patterns to boost the tail samples modeling. To deal with the afore mentioned problem, the features of trajectories within the same pattern cluster are pulled together, while the features from different pattern clusters will be pushed apart. Moreover, a more flexible network structure of the decoder is introduced to exploit the shaped feature embedding space with different pattern clus- ters. Our contribution can be summarized as follows: • We propose a future enhanced contrastive feature learning framework for long-tailed trajectory predic- tion, which can better distinguish tail patterns from head patterns, and the different patterns are repre- sented by different cluster prototypes to enhance the modeling of the tailed data. • We propose a distribution-aware hyper predictor, aim- ing at providing separated decoder parameters for tra- jectory inputs with different patterns. • Experimental results show that our proposed frame- work can outperform start-of-the-art methods.
Wang_MetaMix_Towards_Corruption-Robust_Continual_Learning_With_Temporally_Self-Adaptive_Data_Transformation_CVPR_2023	Abstract Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnera- ble to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety- critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption ro- bustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the general- ization performance against data corruptions during train- ing. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive exper- iments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines.	1. Introduction Humans constantly acquire new information throughout their lifespan and easily recognize information shifts such as structure and style variations in images. Continual learn- ing (CL) aims at imitating human’s ability to learn from non-stationary data distributions without forgetting the previ- ously learned knowledge. The past few years have witnessed rapid progress in CL research [1, 30, 31, 36, 45]. Despite the success, existing CL systems overlook the robustness *Corresponding authoragainst unforeseen data shifts during testing. They assume that training and test images for each task follow the same distribution. However, as data distributions evolve and new scenarios occur, test images often encounter various cor- ruptions such as snow, blur, pixelation, and combinations, resulting in a shifted distribution from the training set. For example, Figure 1 shows various corruptions applied on one image from Split-miniImageNet. Data corruption can drastically impair the performance of existing image recognition systems. A recent study [21] shows that classification accuracy of various architectures has dropped significantly on the ImageNet test set with some simple and natural corruptions. Our empirical evaluation results show that state-of-the-art CL models are even more vulnerable to these corruptions during testing. For example, the accuracy of DER++ [5] for task-continual learning de- creases from 93.9% to 50.5% on split-CIFAR10; accuracy drops to 10.6 % from 75.6 % on split-CIFAR100; accuracy decreases to 9.8% from 61.3% on split-miniImageNet by applying those common corruptions on test data of each CL task. This severe issue makes existing CL models highly unreliable in safety-critical applications. Thus, improving the robustness of CL models to foster their trustworthiness when deployed in real-world scenarios is essential. Training a CL model robust to various corruptions is diffi- cult due to the following challenges. 1) Unseen corruptions could perturb the test set far beyond those encountered dur- ing training. A model that naively augments training images with seen corruptions cannot generalize to the new ones during testing. Also, it is unrealistic to enumerate all possi- ble corruptions during training since there are infinite types of corruptions and their combinations. 2) With the ever- evolving data distributions in CL, an effective augmentation strategy learned on previous tasks may gradually become less effective because the optimal augmentation strategies are task-dependent and dynamically change over tasks [11]. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 24521 (a) Original data (b) Brightness (c) Contrast (d) Defocus Blur (e) Elastic (f) Fog Figure 1. Visualization of five types of different corruption operations on the testing images of Split-miniImageNet. Although we can adopt a memory buffer to store data from previous tasks, augmenting and replaying them at later train- ing iterations, the augmented memory may gradually be- come less effective as the model could memorize the stored information after replay runs. Recent approaches, such as Augmix [22] composes and combines multiple pre-defined augmentation operations with different depths and widths, show efficacy in improving robustness under traditional su- pervised classification tasks. However, these approaches are not directly applicable to corruption-robust CL since they only use a fixed random augmentation strategy that is often not optimal for non-stationary data distributions in CL. To address these unique challenges of corruption- robustness in CL, we propose a temporally self-adaptive Augmix within a meta-learning framework, named MetaMix. It adaptively augments the memory buffer data and the cur- rently received new data by learning to mix the augmentation operations tailored to the evolving data distributions. In par- ticular, our automatic self-adaptive MetaMix is a bi-level optimization, simulating the evaluation process on unseen corruptions. We randomly divide the training augmentation operations into pseudo-seen and pseudo-unseen operations at each CL step. The lower-level optimization is to optimize the model performance on the pseudo-seen operations; the upper-level optimization is to optimize the generalization of the pseudo-unseen operations. The augmentation strategy is governed by an LSTM, which inputs context information and outputs the corresponding mixing parameters for the augmentations. The proposed MetaMix ensures the augmen- tation strategy automatically adapts to non-stationary data distribution. Furthermore, the objective is to optimize the performance of the pseudo-unseen corruption operations, which aligns with our goal during testing and encourages the generalization to unseen corruptions. To evaluate the corruption robustness of existing and the proposed methods, we propose a new challenging benchmark where various corruptions perturb the testing data of each CL task. To facilitate future research, we construct several new datasets, including split-CIFAR-10-C, split-CIFAR-100-C, and split-miniImageNet-C. Extensive experiments on the constructed benchmarks demonstrate the effectiveness of our proposed MetaMix approach compared with several SOTA data-augmentation approaches adapted for CL. We summarize our contributions as follows:•To our best knowledge, we are the first to study the corruption-robustness of CL methods. Accordingly, we propose the first set of novel benchmarks for evaluating the corruption-robustness of existing CL methods and moving towards trustworthy CL. •We propose a self-adaptive augmentation method, MetaMix, by learning to mix and augment the training data of each CL task to achieve corruption-robustness on unseen corruptions for each CL task. •Our method is versatile and can be seamlessly inte- grated with existing CL methods. Extensive experi- ments with both task/class continual learning demon- strate the effectiveness of MetaMix.
Wang_Accelerating_Vision-Language_Pretraining_With_Free_Language_Modeling_CVPR_2023	Abstract The state of the arts in vision-language pretraining (VLP) achieves exemplary performance but suffers from high training costs resulting from slow convergence and long training time, especially on large-scale web datasets. An essential obstacle to training efficiency lies in the entan- gled prediction rate (percentage of tokens for reconstruc- tion) and corruption rate (percentage of corrupted tokens) in masked language modeling (MLM), that is, a proper cor- ruption rate is achieved at the cost of a large portion of output tokens being excluded from prediction loss. To ac- celerate the convergence of VLP , we propose a new pre- training task, namely, free language modeling (FLM), that enables a 100% prediction rate with arbitrary corruption rates. FLM successfully frees the prediction rate from the tie-up with the corruption rate while allowing the corrup- tion spans to be customized for each token to be predicted. FLM-trained models are encouraged to learn better and faster given the same GPU time by exploiting bidirectional contexts more flexibly. Extensive experiments show FLM could achieve an impressive 2.5×pretraining time reduc- tion in comparison to the MLM-based methods, while keep- ing competitive performance on both vision-language un- derstanding and generation tasks. Code will be public athttps://github.com/TencentARC/FLM .	1. Introduction Vision-language pretraining (VLP) has recently demon- strated impressive performance on a handful of vision- language tasks [7,10,14,18,19,22], e.g., visual question an- swering, cross-modal retrieval, and image captioning. Sev- eral factors are responsible for the success: the availabil- ity of large-scale image-text datasets collected from the web [30], high-capacity model architectures like Trans- †Work done during internship in ARC Lab, Tencent PCG. ∗Corresponding author 69.7375.7776.2(7 epoch)75.676.8977.8478.1178.4378.2978.6(41 epoch) 73.9477.2577.5278.63(8 epoch) 697173757779 0102030405060NLVR2 Performance (%)Pretraining Time (GPU Days)ARMLMFLM 70717273747576777879 1.21.41.61.82.02.22.4 020406080100NLVR2 PerformanceValidation MLM Loss Iterations (k)MLM (corr. rate=0.4, pred. rate=0.4)MLM (corr. rate=0.4, pred. rate=0.2)MLM (corr. rate=0.4, pred. rate=0.1) 70717273747576777879 1.21.41.61.82.02.22.4 020406080100NLVR2 PerformanceValidation MLM Loss Iterations (k)MLM (cr=0.4, pr=0.4)MLM (cr=0.4, pr=0.2)MLM (cr=0.4, pr=0.1) 70717273747576777879 1.21.41.61.82.02.22.4 020406080100NLVR2 PerformanceValidation MLM Loss Iterations (k)MLM (cr=0.4, pr=0.4)MLM (cr=0.4, pr=0.2)MLM (cr=0.4, pr=0.1)70717273747576777879 1.21.41.61.82.02.22.4 020406080100NLVR2 PerformanceValidation MLM Loss Iterations (k)MLM (cr=0.4, pr=0.4)MLM (cr=0.4, pr=0.2)MLM (cr=0.4, pr=0.1)Figure 1. (a) Large prediction rate accelerates training. Given a fixed corruption rate, we vary the prediction rate by randomly selecting a subset of output tokens for prediction loss. The learn- ing rate schedule follows METER [10]. (b) The proposed FLM achieves competitive performance compared with MLM mean- while significantly accelerating the pretraining stage. The down- stream performance on NLVR2[32] is reported. We show accu- racy curves before convergence for better visualization. former [34], and effective pretraining objectives for cross- modal learning. One of the dominant pretraining objectives is masked language modeling (MLM), which was first introduced in natural language processing [9] and has been applied to vision-language areas in recent years [19]. MLM is a generative pretraining task designed to reconstruct a few (usually 40% for VLP) masked text tokens via reasoning This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23161 among the context of the remaining texts and the paired image. While effective in capturing cross-modal interac- tions, MLM-based methods [7,15,21] suffer from slow con- vergence and long training time, especially for large-scale models and noisy web data. We argue that the limited prediction rate in MLM im- pedes the convergence speed of pretraining, since a large portion of tokens accompanied by corruption are excluded from prediction loss. As shown in Fig. 1 (top), under the same corruption rate , a larger prediction rate for MLM results in faster convergence of validation loss and down- stream performance. It is intuitive to set a prediction rate of 100% to fully exploit text tokens. However, a paradox emerges where a large prediction rate can only be achieved with a greater corruption rate in MLM, but an extremely large corruption rate leads to an extremely tough pretrain- ing task that may cause training collapse. Autoregressive language modeling (AR) provides a workable solution to enable a 100% prediction rate. It pre- dicts the next token given the observation of previous to- kens. As shown in Fig. 1 (bottom), AR performs favor- ably in training efficiency against MLM, i.e., 6.1×speed-up for convergence. However, the converged performance by AR is, unfortunately, much inferior to MLM. It is probably caused by the sub-optimal unidirectional corruption pattern, which is insufficient for downstream understanding tasks that usually rely on bidirectional contexts. A natural question arises, can we accelerate the conver- gence of VLP by predicting 100% tokens like AR mean- while achieving competitive performance with MLM? To- wards this end, we introduce a new pretraining task, dubbed free language modeling (FLM), for VLP, that enjoys an ex- treme 100% prediction rate and flexible bidirectional con- textualized representations. We for the first time break up the entanglement between corruption and prediction rates, making the two factors freely determined. Furthermore, for each output token to be predicted, we allow independent and arbitrary-length spans (from one to 100% tokens) as cor- rupted connections. Rather than the suffix-like corruption pattern as in AR (as well as PrefixLM [37]), the corruption span of FLM is primarily distributed in the middle of the sequence, establishing a flexible perception of bidirectional contexts for better adaptation to VL understanding tasks. The comparison between different pretraining objectives is illustrated in Fig. 2. To perform VLP with FLM, we propose an encode- corrupt-predict framework, which performs feature encod- ing once and reconstructs several corrupted versions of the text sequence in parallel. In the encoding step, bidirec- tional representations are achieved by learning forward and reverse unidirectional representations respectively, the or- der of which is manipulated by (reverse) casual masks in the same text Transformer. Subsequently, we ensure a100% prediction rate by customizing corruption-prediction tasks for predicting each input token. In each corruption- prediction task, a span of corruption is randomly sampled and attached to the encoded sequence, followed by a re- constructor to solve the prediction task by reasoning among the remaining contexts. Unlike previous works ( e.g., MLM, AR) that adopt pre-encoding corruption, we inject corrup- tions after one-time feature encoding, encouraging flexible corruption patterns and efficient parallel prediction. Our contributions are three-fold. (1) A novel pretraining objective for VLP, namely, free language modeling (FLM), is proposed to free the prediction rate from the constraints of corruption rate, enabling an appealing 100% prediction rate for accelerating convergence speed during pretraining. (2) An encode-corrupt-predict framework built upon FLM ob- jective is proposed, allowing efficient and effective learning of a set of prediction tasks by merely conducting feature en- coding once. (3) Extensive experiments on VQA, NLVR2, image captioning, and image-text retrieval demonstrate the effectiveness of our FLM, where comparable performances to MLM are achieved with less than 50% pretraining time.
Wang_Turning_Strengths_Into_Weaknesses_A_Certified_Robustness_Inspired_Attack_Framework_CVPR_2023	Abstract Graph neural networks (GNNs) have achieved state-of- the-art performance in many graph learning tasks. How- ever, recent studies show that GNNs are vulnerable to both test-time evasion and training-time poisoning attacks that perturb the graph structure. While existing attack methods have shown promising attack performance, we would like to design an attack framework to further enhance the perfor- mance. In particular, our attack framework is inspired by certiﬁed robustness, which was originally used by defend- ersto defend against adversarial attacks. We are the ﬁrst, from the attacker perspective, to leverage its properties to better attack GNNs. Speciﬁcally, we ﬁrst derive nodes’ cer- tiﬁed perturbation sizes against graph evasion and poison- ing attacks based on randomized smoothing, respectively. A larger certiﬁed perturbation size of a node indicates this node is theoretically more robust to graph perturbations. Such a property motivates us to focus more on nodes with smaller certiﬁed perturbation sizes, as they are easier to be attacked after graph perturbations. Accordingly, we design a certiﬁed robustness inspired attack loss, when incorpo- rated into (any) existing attacks, produces our certiﬁed ro- bustness inspired attack counterpart. We apply our frame- work to the existing attacks and results show it can signiﬁ- cantly enhance the existing base attacks’ performance.	1. Introduction Learning with graphs, such as social networks, citation networks, chemical networks, has attracted signiﬁcant at- tention recently. Among many methods, graph neural net- works (GNNs) [ 14,33,38,41,44] have achieved state-of-the- art performance in graph related tasks such as node classi- ﬁcation, graph classiﬁcation, and link prediction. However, recent studies [ 8,19,20,23,30,34,36,37,39,40,50,51] show that GNNs are vulnerable to both test-time graph evasion *Corresponding authorsattacks and training-time graph poisoning attacks1. Take GNNs for node classiﬁcation as an instance, graph eva- sion attacks mean that, given a learnt GNN model and a (clean) graph, an attacker carefully perturbs the graph struc- ture (i.e., inject new edges to or remove the existing edges from the graph) such that as many testing nodes as possi- ble are misclassiﬁed o by the GNN model. Whereas, graph poisoning attacks mean that, given a GNN algorithm and a graph, an attacker carefully perturbs the graph structure in the training phase, such that the learnt GNN model misclas- siﬁes as many testing nodes as possible in the testing phase. While existing methods have shown promising attack per- formance, we want to ask: Can we design a general attack framework that can further enhance both the existing graph evasion and poisoning attacks to GNNs? The answer is yes. We design an attack framework inspired by certiﬁed ro- bustness. Certiﬁed robustness was originally used by de- fenders to guarantee the robustness of classiﬁcation mod- els against evasion attacks. Generally speaking, a testing example (e.g., an image or a node) with a better certiﬁed robustness guarantee indicates this example is theoretically more robust to adversarial (e.g., pixel or graph) perturba- tions. While certiﬁed robustness is mainly derived for do- ing the good, attackers , on the other hand, can also leverage its property to do the bad. For instance, when an attacker knows the certiﬁed robustness of nodes in a graph, he can base on nodes’ certiﬁed robustness to reversely reveal the vulnerable region of the graph and leverage this vulnerabil- ity to design better attacks. We are inspired by such prop- erty of certiﬁed robustness and design the ﬁrst certiﬁed ro- bustness inspired attacks to GNNs. Our attack framework consists of three parts: i) In- spired by the state-of-the-art randomized smoothing based certiﬁed robustness against evasion attacks to image mod- els [7,28] and GNN models [ 35], we ﬁrst propose to gen- eralize randomized smoothing and derive the node’s cer- 1We mainly consider the graph structure attack in the paper, as it is more effective than the feature attack. However, our attack framework can be easily extended to the feature attack. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 16394 tiﬁed perturbation size against graph poisoning attacks to GNNs. Particularly, a larger certiﬁed perturbation size of a node indicates this node is theoretically more robust to ad- versarial graph perturbations. In other words, an attacker needs to perturb more edges during the training phase in order to make this node wrongly predicted by the learnt GNN model. This property inspires us to focus more on disrupting nodes with relatively smaller certiﬁed perturba- tion sizes under a given perturbation budget. ii) We design a certiﬁed robustness inspired attack loss. Speciﬁcally, we modify the classic node-wise loss by assigning each node a weight based on its certiﬁed perturbation size—A node with a larger/smaller certiﬁed perturbation size will be as- signed a smaller/larger weight. In doing so, losses for nodes with smaller certiﬁed perturbation sizes will be enlarged, and most of the perturbation budget will be automatically allocated to perturb these nodes. Thus, more nodes will be misclassiﬁed with the given perturbation budget. iii) We design the certiﬁed robustness inspired attack framework to generate adversarial graph perturbations to GNNs, based on our certiﬁed robustness inspired attack loss. We emphasize that, as our new attack loss only modiﬁes the existing attack loss with certiﬁed perturbation size deﬁned node weights, any existing graph evasion or poisoning attack method can be used as the base attack in our framework. We apply our certiﬁed robustness inspired attack frame- work to the state-of-the-art graph evasion and poisoning attacks [ 40,51] to GNNs. Evaluation results on multiple benchmark datasets show our attack framework can sub- stantially enhance the attack performance of the base at- tacks. Our contributions are as follows: •We propose a certiﬁed robustness inspired attack frame- work to GNNs. Our framework can be plugged into any existing graph evasion and poisoning attacks. •To our best knowledge, we are the ﬁrst work to use certi- ﬁed robustness for an attack purpose. •Evaluation results validate the effectiveness of our attack framework when applied to the existing attacks to GNNs.
Wu_MagicPony_Learning_Articulated_3D_Animals_in_the_Wild_CVPR_2023	Abstract We consider the problem of predicting the 3D shape, ar- ticulation, viewpoint, texture, and lighting of an articulated animal like a horse given a single test image as input. We present a new method, dubbed MagicPony, that learns this predictor purely from in-the-wild single-view images of the object category, with minimal assumptions about the topol- ogy of deformation. At its core is an implicit-explicit repre- sentation of articulated shape and appearance, combining the strengths of neural fields and meshes. In order to help the model understand an object’s shape and pose, we distil the knowledge captured by an off-the-shelf self-supervised vision transformer and fuse it into the 3D model. To over- come local optima in viewpoint estimation, we further in- troduce a new viewpoint sampling scheme that comes at no additional training cost. MagicPony outperforms prior work on this challenging task and demonstrates excellent generalisation in reconstructing art, despite the fact that it is only trained on real images. The code can be found on the project page at https://3dmagicpony.github.io/ .	1. Introduction Reconstructing the 3D shape of an object from a sin- gle image of it requires knowing a priori what are the possible shapes and appearances of the object. Learning *Equal contribution.such a prior usually requires ad-hoc data acquisition se- tups [2, 20, 35, 36], involving at least multiple cameras, and often laser scanners, domes and other hardware, not to men- tion significant manual effort. This is viable for certain types of objects such as humans that are of particular in- terest in applications, but it is unlikely to scale to the long tail of objects that can appear in natural images. The alter- native is to learn a 3D prior from 2D images only, which are available in abundance. However, this prior is highly com- plex and must be learned while using it to reconstruct in 3D the 2D training data, which is a major challenge. In this paper, we propose MagicPony , a novel approach to learning 3D models of articulated object categories such as horses and birds with only single-view input images for training. We leverage recent progress in unsupervised rep- resentation learning, unsupervised image matching, effi- cient implicit-explicit shape representations and neural ren- dering, and devise a new auto-encoder architecture that re- constructs the 3D shape, articulation and texture of each ob- ject instance from a single image. For training, we only require a 2D segmenter for the object category and a de- scription of the topology and symmetry of its 3D skeleton (i.e., the number and connectivity of bones). We do not require apriori knowledge of the objects’ 3D shapes, key- points, viewpoints, or of any other 2D or 3D cue which are often used in prior work [14, 21, 22, 31]. From this, we learn a function that, at test time, can estimate the shape and texture of a new object from a single image, in a feed- forward manner. The function exhibits remarkable gener- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8792 alisation properties, including reconstructing objects in ab- stract drawings , despite being trained on real images only. In order to learn such disentangled 3D representations simply from raw images, MagicPony addresses a few key challenges. The first challenge is viewpoint estimation. Be- fore the 3D model is available, it is very difficult to assign a viewpoint to the objects in the training images. To tackle this problem, prior works have often assumed additional cues such as 2D keypoint correspondences [21, 22, 31]. In- stead, we avoid requiring additional keypoint supervision and implicitly infer noisy correspondences by fusing into the 3D model knowledge distilled from DINO-ViT [5], a self-supervised visual transformer network (ViT) [28]. We also develop a new efficient disambiguation scheme that explores multiple viewpoint assignment hypotheses at es- sentially no cost, avoiding local optima that are caused by greedily matching the noisy 2D correspondences. The second challenge is how to represent the 3D shape, appearance and deformations of the object. Most prior works have used textured meshes [13,14,21,31,32,67,70], but these are difficult to optimise from scratch, leading to problems that often require ad-hoc heuristics such as re- meshing [13, 70]. The other and increasingly popular ap- proach is to use a volumetric representation such as a neu- ral radiance field [3, 40, 52, 63], which can model complex shapes, including manipulating their topology during train- ing. However, this modelling freedom comes at the cost of over-parametrisation, which is particularly problematic in monocular reconstruction and often leads to meaning- less short-cut solutions [63]. Furthermore, modelling ar- ticulation with a volumetric representation is difficult. A posing transformation is only defined for the object’s sur- face and interior and is more easily expressed from the canonical/pose-free space to the posed space. However, rendering a radiance field requires transforming 3D points off the object surface and in the direction opposite to the posing transformation, which is hard [8]. We address these issues by using a hybrid volumetric- mesh representation based on DMTet [42, 56]. Shape and appearance are defined volumetrically in canonical space, but a mesh is extracted on the fly for posing and rendering. This sidesteps the challenges of using neural rendering di- rectly while retaining most of its advantages and enabling the use of powerful shape regularisers. To summarise, we make the following contributions : (1) A new 3D object learning framework that combines re- cent advances in unsupervised learning, 3D representations and neural rendering, achieving better reconstruction results with less supervision; (2) An effective mechanism for fus- ing self-supervised features from DINO-ViT into the 3D model as a form of self-supervision; and (3) an efficient multi-hypothesis viewpoint prediction scheme that avoids local optima in reconstruction with no additional cost.Table 1. Related Work Overview on Weakly-supervised Learn- ing of 3D Objects. Annotations: template shape, ὏7viewpoint, ὑ12D keypoint, /ctobject mask,/frardoptical flow,video,1coarse template shape from keypoints,2camera estimated from keypoints using SfM,3outputs texture flow,4shape bases initialised from CMR.†UMR relies on part segmentations from SCOPS [18]. Supervision Output Method ὏7 ὑ1/ct/frard 3D 2.5D Motion View Texture Unsup3D [69] ✓ ✓ ✓ CSM [30] ✓ ✓ ✓ A-CSM [29] ✓ ✓ ✓ ✓ CMR [21] ( ✓)1(✓)2✓ ✓ ✓ ✓ (✓)3 U-CMR [14] ✓ ✓ ✓ ✓ ✓ UMR†[32] ✓ ✓ ✓ (✓)3 ACMR [31] ( ✓)4(✓)2✓ ✓ ✓ ✓ (✓)3 DOVE [67] ✓ ✓ ✓ ✓ ✓ ✓ ✓ Ours ✓ ✓ ✓ ✓ ✓ We compare our method to prior work on several chal- lenging articulated animal categories and show that our method obtains significantly better quantitative and quali- tative results while using significantly less supervision, and also demonstrate generalisation to abstract drawings.
Wang_Towards_Domain_Generalization_for_Multi-View_3D_Object_Detection_in_Bird-Eye-View_CVPR_2023	Abstract Multi-view 3D object detection (MV3D-Det) in Bird- Eye-View (BEV) has drawn extensive attention due to its low cost and high efficiency. Although new algorithms for camera-only 3D object detection have been continu- ously proposed, most of them may risk drastic performance degradation when the domain of input images differs from that of training. In this paper, we first analyze the causes of the domain gap for the MV3D-Det task. Based on the covariate shift assumption, we find that the gap mainly at- tributes to the feature distribution of BEV , which is deter- mined by the quality of both depth estimation and 2D im- age’s feature representation. To acquire a robust depth pre- diction, we propose to decouple the depth estimation from the intrinsic parameters of the camera (i.e. the focal length) through converting the prediction of metric depth to that of scale-invariant depth and perform dynamic perspective augmentation to increase the diversity of the extrinsic pa- rameters (i.e. the camera poses) by utilizing homography. Moreover, we modify the focal length values to create mul- tiple pseudo-domains and construct an adversarial train- ing loss to encourage the feature representation to be more domain-agnostic. Without bells and whistles, our approach, namely DG-BEV , successfully alleviates the performance drop on the unseen target domain without impairing the accuracy of the source domain. Extensive experiments on Waymo, nuScenes, and Lyft, demonstrate the generalization and effectiveness of our approach.	1. Introduction 3D object detection, aiming at localizing objects in the 3D space, is critical for various applications such as au- *Shuo Wang and Xinhai Zhao contributed equally. This work was done when Shuo Wang was an intern at Huawei Noah’s Ark Lab. †Corresponding author. (a) Baseline (b) DG-BEV Figure 1. Qualitative comparisons between BEVDepth and the proposed DG-BEV . The red and blue bounding boxes represent ground truth and detected results on the target domain respectively. Depth-shift is shown in green arrows. Our approach can detect correct 3D results on unknown domains. tonomous driving [6, 38], robotic navigation [2], and vir- tual reality [32], etc. Despite the remarkable progress of LiDAR-based methods [17,30,33], camera-based 3D object detection in Bird-Eye-View (BEV) [14, 19, 21] has drawn increasing attention in recent years due to its rich semantic information and low cost for deployment. However, most of the detectors assume that the training and testing data are obtained in the same domain which may be hardly guaranteed in realistic scenarios. Thus, tremen- dous performance degradation will appear when the domain of the input image shifts. For example, nuScenes [3] and This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13333 Figure 2. Illustration of the difficulty in estimating depth based on cameras with different focal length. O1andO2are the optical centers of two cameras and C is the object being photographed. A and B denote the imaging planes of the two cameras respectively and the red parts show the size of the same object in their corre- sponding image planes. Waymo [34] are two popular benchmarks for 3D object de- tection and their data collection devices are not identical, i.e., both of the intrinsic and extrinsic parameters are differ- ent. Empirical results presented in Fig. 1 show that detec- tors trained on nuScenes have location bias when predicting objects on the Waymo dataset. Domain Generalization (DG) [8, 18, 26], aiming to learn a model that generalizes well on unseen target domains, can be a plausible solution to alleviate the bias mentioned above. In the literature, DG has been widely explored for 2D vision tasks, e.g., image recognition [7, 16], object de- tection [31, 44], and semantic segmentation [27, 41]. How- ever, most of these works are designed for the case where there are multiple source domains available which are ob- viously infeasible due to the diversity of the real world in autonomous driving scenarios. Alternatively, one recent work [39] proposed to study the single-domain generaliza- tion for LiDAR-based detection. However, it is not tractable to directly adapt this method to solve the camera-based de- tection task due to the fundamental differences between the characteristics of points and images. Therefore, developing a general domain generalization framework for MV3D-Det is still highly desirable. In this paper, we theoretically analyze the causes of the domain gap for MV3D-Det. Based on the covariate shift assumption [4], we find that such a gap mainly attributes to the feature distribution of BEV , which is determined by the depth estimation and 2D image feature jointly. Based on this, we propose DG-BEV , a domain generalization method for MV3D-Det in BEV . Specifically, we first conduct a thor- ough analysis of why the estimated depth becomes inaccu- rate when the domain shifts and find the key factor lies in that intrinsic parameters of cameras used in various domains are hardly guaranteed to be identical (please refer to Fig. 2 for a better understanding). To alleviate this issue, we pro- pose to decouple the depth estimation from the intrinsic pa- rameters by converting the prediction of metric depth to that of scale-invariant depth. On the other hand, extrinsic param-eters of cameras ( e.g. camera poses) also play an important role in camera-based depth estimation, which is often ig- nored in previous works. Instead, we introduce homogra- phy learning to dynamically augment the image perspec- tives by simultaneously adjusting the imagery data and the camera pose. Moreover, since domain-agnostic feature representations are favored for better generalization, we propose to build up multiple pseudo-domains by modifying the focal length values of camera intrinsic parameters in the source domain and construct an adversarial training loss to further enhance the quality of feature representations. In summary, the main contributions of this paper are: •We present a theoretical analysis on the causes of the domain gap in MV3D-Det. Based on the covariate shift assumption, we find the gap lies in the feature distribution of BEV , which is determined by the depth estimation and 2D image feature jointly. •We propose DG-BEV , a domain generalization method to alleviate the domain gap from both of the two per- spectives mentioned above. •Extensive experiments on various public datasets, in- cluding Waymo, nuScenes, and Lyft, demonstrate the generalization and effectiveness of our approach. •To the best of our knowledge, this is the first systematic study to explore a domain generalization method for multi-view 3D object detectors.
Wang_Score_Jacobian_Chaining_Lifting_Pretrained_2D_Diffusion_Models_for_3D_CVPR_2023	Abstract A diffusion model learns to predict a vector field of gradi- ents. We propose to apply chain rule on the learned gradients, and back-propagate the score of a diffusion model through the Jacobian of a differentiable renderer, which we instan- tiate to be a voxel radiance field. This setup aggregates 2D scores at multiple camera viewpoints into a 3D score, and re- purposes a pretrained 2D model for 3D data generation. We identify a technical challenge of distribution mismatch that arises in this application, and propose a novel estimation mechanism to resolve it. We run our algorithm on several off- the-shelf diffusion image generative models, including the recently released Stable Diffusion trained on the large-scale LAION 5B dataset.	1. Introduction We introduce a method that converts a pretrained 2D diffusion generative model on images into a 3D generative model of radiance fields, without requiring access to any 3D data. The key insight is to interpret diffusion models as * Equal contribution.learned predictors of a gradient field, often referred to as the score function of the data log-likelihood. We apply the chain rule on the estimated score, hence the name Score Jacobian Chaining (SJC). Following Hyvärinen [16], the score is defined as the gradient of the log-density function with respect to the data (rather than parameter). Diffusion models of various fam- ilies [ 13,49,50,52] can all be interpreted [ 20,23,52] as modeling ∇xlogpσ(x)i.e. the denoising score at noise levelσ. For readability, we refer to the denoising score as the score. Generating a sample from a diffusion model involves repeated evaluations of the score function from large to small σlevel, so that a sample xgradually moves closer to the data manifold. It can be loosely interpreted as gradient descent, with precise control on the step sizes so that data distribution evolves to match the annealed σlevel (ancestral sampler [ 13], SDE and probability-flow ODE [ 52], etc.). While there are other perspectives to a diffusion model [ 13,49], here we are primarily motivated from the viewpoint that diffusion models produce a gradient field. A natural question to ask is whether the chain rule can be applied to the learned gradients. Consider a diffusion model on images. An image xmay be parameterized by some This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12619 function fwith parameters θ,i.e.,x=f(θ). Applying the chain rule through the Jacobian∂x ∂θconverts a gradient on image xinto a gradient on the parameter θ. There are many potential use cases for pairing a pretrained diffusion model with different choices of f. In this work we are interested in exploring the connection between 3D and multiview 2D by choosing fto be a differentiable renderer, thus creating a 3D generative model using only pretrained 2D resources. Many prior works [ 2,58,60] perform 3D generative mod- eling by training on 3D datasets [ 5,24,54,59]. This ap- proach is often as challenging as it is format-ambiguous. In addition to the high data acquisition cost of 3D assets [ 9], there is no universal data format: point clouds, meshes, volu- metric radiance field, etc, all have computational trade-offs. What is common to these 3D assets is that they can be ren- dered into 2D images. An inverse rendering system, or a differentiable renderer [ 25,27,30,34,39], provides access to the Jacobian Jπ≜∂xπ ∂θof a rendered image xπat camera viewpoint πwith respect to the underlying 3D parameteriza- tionθ. Our method uses differentiable rendering to aggregate 2D image gradients over multiple viewpoints into a 3D asset gradient, and lifts a generative model from 2D to 3D. We parameterize a 3D asset θas a radiance field stored on voxels and choose fto be the volume rendering function. A key technical challenge is that computing the 2D score by directly evaluating a diffusion model on a rendered image xπleads to an out-of-distribution (OOD) problem. Gener- ally, diffusion models are trained as denoisers and have only seen noisy inputs during training. On the other hand, our method requires evaluating the denoiser on non-noisy ren- dered images from a 3D asset during optimization, and it leads to the OOD problem. To address the issue, we propose Perturb-and-Average Scoring , an approach to estimate the score for non-noisy images. Empirically, we first validate the effectiveness of Perturb- and-Average Scoring at solving the OOD problem and ex- plore the hyperparameter choices on a simple 2D image can- vas. Here we identify open problems on using unconditioned diffusion models trained on FFHQ and LSUN Bedroom. Next, we use Stable Diffusion, a model pretrained on the web-scale LAION dataset to perform SJC for 3D generation, as shown in Fig. 1. Our contributions are as follows: •We propose a method for lifting a 2D diffusion model to 3D via an application of the chain rule. •We illustrate the challenge of OOD when using a pretrained denoiser and propose Perturb-and-Average Scoring to resolve it. •We point out the subtleties and open problems on ap- plying Perturb-and-Average Scoring as gradient for optimization. •We demonstrate the effectiveness of SJC for the task of 3D text-driven generation.
Xu_Iterative_Geometry_Encoding_Volume_for_Stereo_Matching_CVPR_2023	Abstract Recurrent All-Pairs Field Transforms (RAFT) has shown great potentials in matching tasks. However, all-pairs cor- relations lack non-local geometry knowledge and have dif- ficulties tackling local ambiguities in ill-posed regions. In this paper, we propose Iterative Geometry Encoding Volume (IGEV-Stereo), a new deep network architecture for stereo matching. The proposed IGEV-Stereo builds a combined geometry encoding volume that encodes geometry and con- text information as well as local matching details, and itera- tively indexes it to update the disparity map. To speed up the convergence, we exploit GEV to regress an accurate starting point for ConvGRUs iterations. Our IGEV-Stereo ranks 1st on KITTI 2015 and 2012 (Reflective) among all published methods and is the fastest among the top 10 methods. In addition, IGEV-Stereo has strong cross-dataset generaliza- tion as well as high inference efficiency. We also extend our IGEV to multi-view stereo (MVS), i.e. IGEV-MVS, which achieves competitive accuracy on DTU benchmark. Code is available at https://github.com/gangweiX/IGEV.	1. Introduction Inferring 3D scene geometry from captured images is a fundamental task in computer vision and graphics with applications ranging from 3D reconstruction, robotics and autonomous driving. Stereo matching which aims to re- construct dense 3D representations from two images with calibrated cameras is a key technique for reconstructing 3D scene geometry. Many learning-based stereo methods [5, 17, 24, 47, 48] have been proposed in the literature. The popular repre- sentative is PSMNet [5] which apply a 3D convolutional encoder-decoder to aggregate and regularize a 4D cost vol- ume and then use soft argmin to regress the disparity map from the regularized cost volume. Such 4D cost volume filtering-based methods can effectively explore stereo ge- ometry information and achieve impressive performance on †Corresponding author. (a) (b)Figure 1. (a) Comparison with state-of-the-art stereo methods [9, 21, 25, 43, 47, 59] on KITTI 2012 and 2015 leaderboards. (b) Performance comparison with RAFT-Stereo [24] on Scene Flow test set as the number of iterations changes. several benchmarks. However, it usually demands a large amount of 3D convolutions for cost aggregation and regu- larization, and in turn yield high computational and memory costs. As a result, it can hardly be applied to high-resolution images and/or large-scale scenes. Recently, iterative optimization-based methods [21, 24, 30, 39, 43] have exhibited attractive performance on both high resolution images and standard benchmarks. Different from existing methods, iterative methods bypass the com- putationally expensive cost aggregation operations and pro- gressively update the disparity map by repeatedly fetching information from a high-resolution 4D cost volume. Such solution enables the direct usage of high-resolution cost vol- ume and hence is applicable to high-resolution images. For instance, RAFT-Stereo [24] exploits a multi-level Convo- lutional Gated Recurrent Units (ConvGRUs) [10] to recur- rently update the disparity field using local cost values re- trieved from all-pairs correlations (APC). However, without cost aggregation the original cost vol- ume lacks non-local geometry and context information (see Fig. 2 (b)). As a result, existing iterative methods have difficulties tackling local ambiguities in ill-posed regions, such as occlusions, texture-less regions and repetitive struc- tures. Even though, the ConvGRU-based updater can im- prove the predicted disparities by incorporating context and geometry information from context features and hidden lay- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21919 (a) Left image (b) Disparity from APC (c) Disparity from GEV (d) Final disparity Figure 2. (a) Input images from KITTI 2015. Illustration of (b) disparity regressed from All-pairs Correlations (APC) in RAFT-Stereo [24], (c) disparity regressed from our Geometry Encoding V olume (GEV), (d) our final disparity. The APC lacks non-local geometry knowledge and thus has difficulties tackling local ambiguities in ill-posed region. We take full advantage of cost filtering and iterative optimization: 1) exploiting 3D CNN to filter cost volume and obtain the strong scene representation and the initial disparity with smooth edges, 2) exploiting ConvGRUs to optimize the initial disparity to recover object edges and details. ers, such limitation in the original cost volume greatly lim- its the effectiveness of each iteration and in turn yields a large amount of ConvGRUs iterations for satisfactory per- formance. We claim that cost filtering-based methods and itera- tive optimization-based methods have complementary ad- vantages and limitations. The former can encode sufficient non-local geometry and context information in the cost vol- ume which is essential for disparity prediction in particu- lar in challenging regions. The latter can avoid high com- putational and memory costs for 3D cost aggregation, yet are less capable in ill-posed regions based only on all-pairs correlations. To combine complementary advantages of the two methods, we propose Iterative Geometry Encoding V ol- ume (IGEV-Stereo), a new paradigm for stereo matching (see Fig. 3). To address ambiguities caused by ill-posed regions, we compute a Geometry Encoding V olume (GEV) by aggregating and regularizing a cost volume using an ex- tremely lightweight 3D regularization network. Compared to all-pairs correlations of RAFT-Stereo [24], our GEV en- codes more geometry and context of the scene after aggre- gation, shown in Fig. 2 (c). A potential problem of GEV is that it could suffer from over-smoothing at boundaries and tiny details due to the 3D regularization network. To com- plement local correlations, we combine the GEV and all- pairs correlations to form a Combined Geometry Encoding V olume (CGEV) and input the CGEV into the ConvGRU- based update operator for iterative disparity optimization. Our IGEV-Stereo outperforms RAFT-Stereo in terms of both accuracy and efficiency. The performance gains come from two aspects. First, our CGEV provides more compre- hensive yet concise information for ConvGRUs to update, yielding more effective optimization in each iteration and in turn could significantly reduce the amount of ConvGRUs iterations. As shown in Fig. 1, our method achieves even smaller EPE (i.e., 0.58) using only 3 ConvGRUs iterations (i.e.,100ms totally for inference) than RAFT-Stereo using 32 ConvGRUs iterations (i.e., EPE of 0.61 and 440ms for inference). Second, our method regresses an initial disparity map from the GEV via soft argmin which could providean accurate starting point for the ConvGRU-based update operator, and in turn yield a fast convergence. In compari- son, RAFT-Stereo starts disparity prediction from an initial starting point d0=0, which demands a large number Con- vGRUs iterations to achieve an optimized result. We demonstrate the efficiency and effectiveness of our method on several stereo benchmarks. Our IGEV-Stereo achieves the state-of-the-art EPE of 0.47 on Scene Flow [31] and ranks 1ston KITTI 2015 [32] and 2012 (Re- flective) [15] leaderboards among all the published meth- ods. Regarding the inference speed, our IGEV-Stereo is the fastest among the top 10 methods on KITTI leader- boards. IGEV-Stereo also exhibits better cross-dataset gen- eralization ability than most existing stereo networks. When trained only on synthetic data Scene Flow, our IGEV-Stereo performs very well on real datasets Middlebury [34] and ETH3D [35]. We also extend our IGEV to MVS, i.e. IGEV- MVS, which achieves competitive accuracy on DTU [1].
Wang_AutoRecon_Automated_3D_Object_Discovery_and_Reconstruction_CVPR_2023	Abstract A fully automated object reconstruction pipeline is cru- cial for digital content creation. While the area of 3D recon- struction has witnessed profound developments, the removal of background to obtain a clean object model still relies on different forms of manual labor, such as bounding box labeling, mask annotations, and mesh manipulations. In this paper, we propose a novel framework named AutoRe- con for the automated discovery and reconstruction of an object from multi-view images. We demonstrate that fore- ground objects can be robustly located and segmented from SfM point clouds by leveraging self-supervised 2D vision transformer features. Then, we reconstruct decomposed neu- ral scene representations with dense supervision provided by the decomposed point clouds, resulting in accurate ob- ject reconstruction and segmentation. Experiments on the DTU, BlendedMVS and CO3D-V2 datasets demonstrate the effectiveness and robustness of AutoRecon. The code and supplementary material are available on the project page: https://zju3dv.github.io/autorecon/ .	1. Introduction 3D object reconstruction has long been investigated in computer vision. In this work, we focus on the specific set- ting of reconstructing a salient foreground object from multi- view images and automatically segmenting the object from the background without any annotation, which enables scal- able 3D content creation for VR/AR and may open up the possibility to generate free 2D and 3D object annotations at a large scale for supervised-learning tasks. Traditional multi-view stereo [8, 32] and recent neural scene reconstruction methods [40, 46] have attained impres- sive reconstruction quality. However, these methods cannot identify objects and the reconstructed object models are typ- ically coupled with the surrounding background. A straight- forward solution is utilizing the foreground object masks to The authors are affiliated with the ZJU-SenseTime Joint Lab of 3D Vision. †Corresponding author: Xiaowei Zhou. Coarse SegmentationInput Images Trans. Structure-from-Motion DINO Point CloudForeground Reconstruction Foreground Segmentation SDF Radiance FieldReg.Figure 1. Overview of our fully-automated pipeline and results. Given an object-centric video, we achieve coarse decomposition by segmenting the salient foreground object from a semi-dense SfM point cloud, with pointwise-aggregated 2D DINO features [3]. Then we train a decomposed neural scene representation from multi-view images with the help of coarse decomposition results to reconstruct foreground objects and render multi-view consistent high-quality foreground masks. obtain clean foreground object models. However, accurate 2D object masks are expensive to annotate, and salient ob- ject segmentation techniques [21, 34, 41] generally produce masks with limited granularity, thus degrading the recon- struction quality, especially for objects with thin structures. Recently, some methods [23,30,50] attempt to automatically decompose objects from 3D scenes given minimal human annotations, such as 3D object bounding boxes, scribbles or pixel labels. But the requirement of manual annotations limits the feasibility of more scalable 3D content creation. In this paper, we propose a novel two-stage framework for the fully-automated 3D reconstruction of salient objects, as illustrated in Fig. 1. We first perform coarse decomposition to automatically segment the foreground SfM point cloud, and then reconstruct the foreground object geometry by learning an implicit neural scene representation under explicit super- vision from the coarse decomposition. The key idea of our coarse decomposition is to leverage the semantic features This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21382 provided by a self-supervised 2D Vision Transformer (ViT) [3]. Specifically, we aggregate multi-view ViT features from input images to the SfM point cloud and then segment salient foreground points with a point cloud segmentation Trans- former. To train the Transformer on large-scale unlabeled data, we devise a pseudo-ground-truth generation pipeline based on Normalized Cut [33] and show its ability to pro- duce accurate segmentations and 3D bounding boxes upon training. For object reconstruction, we learn a neural scene representation within the estimated foreground bounding box from multi-view images. Our main idea is to reconstruct a decomposed scene representation with the help of explicit regularization provided by the previously decomposed point cloud. Finally, we can extract a clean object model and obtain high-quality object masks with foreground-only rendering. We conduct experiments on the CO3D [29], Blended- MVS [45], and DTU [12] datasets to validate the effective- ness of the proposed pipeline. The experimental results show that our approach can automatically and robustly recover accurate 3D object models and high-quality segmentation masks from RGB videos, even with cluttered backgrounds. In summary, we make the following contributions: •We propose a fully-automated framework for recon- structing background-free object models from multi- view images without any annotation. •We propose a coarse-to-fine pipeline for scene decom- position by first decomposing the scene in the form of an SfM point cloud, which then guides the decomposi- tion of a neural scene representation. •We propose an SfM point cloud segmentation Trans- former and devise an unsupervised pseudo-ground-truth generation pipeline for its training. •We demonstrate the possibility of automatically creat- ing object datasets with 3D models, 3D bounding boxes, and 2D segmentation masks.
Wei_Joint_Token_Pruning_and_Squeezing_Towards_More_Aggressive_Compression_of_CVPR_2023	Abstract Although vision transformers (ViTs) have shown promis- ing results in various computer vision tasks recently, their high computational cost limits their practical applications. Previous approaches that prune redundant tokens have demonstrated a good trade-off between performance and computation costs. Nevertheless, errors caused by prun- ing strategies can lead to significant information loss. Our quantitative experiments reveal that the impact of pruned tokens on performance should be noticeable. To address this issue, we propose a novel joint Token Pruning & Squeezing module (TPS) for compressing vision transform- ers with higher efficiency. Firstly, TPS adopts pruning to get the reserved and pruned subsets. Secondly, TPS squeezes the information of pruned tokens into partial reserved to- kens via the unidirectional nearest-neighbor matching and similarity-based fusing steps. Compared to state-of-the- art methods, our approach outperforms them under all to- ken pruning intensities. Especially while shrinking DeiT- tiny&small computational budgets to 35%, it improves the accuracy by 1%-6% compared with baselines on ImageNet classification. The proposed method can accelerate the throughput of DeiT-small beyond DeiT-tiny, while its accu- racy surpasses DeiT-tiny by 4.78%. Experiments on various transformers demonstrate the effectiveness of our method, while analysis experiments prove our higher robustness to the errors of the token pruning policy. Code is available at https://github.com/megvii-research/TPS- CVPR2023 .	1. Introduction The transformer architecture has become popular for var- ious natural language processing (NLP) tasks, and its im- proved variants have been adopted for many vision tasks. Vision transformers (ViTs) [5] leverage the long-range de- *The first two authors contributed equally to this work †Corresponding author Input Image Token PruningLabel: Lawn Mower OursLabel: Baseball Prediction: Folding Chair × Prediction: Rugby Ball × Prediction: Lawn Mower √ Prediction: Baseball √Figure 1. Comparisons between token pruning paradigm [25] (the 2nd row) and our joint Token Pruning & Squeezing (the 3rd row). The context information, such as the sod in the examples, is help- ful for prediction but is discarded. Our method remits the informa- tion loss by squeezing the pruned tokens into reserved ones instead of naively dropping them, as indicated by the stacked patches. By this design, we could apply more aggressive token pruning with less performance drop. The example results are from the Ima- geNet1K [4], and we reduce the actual patches grid 14×14to 7×7for visualization clarity. pendencies of self-attention mechanisms to achieve excel- lent performance, often surpassing that of CNNs. In ad- dition to the vanilla ViT architecture, recent studies [17, 31, 33] have explored hybrid ViT designs incorporating convolution layers and multi-scale architectures. Despite their excellent performance, transformers still require rel- atively high computational budgets. This is due to the quadratic computation and memory costs associated with token length. To address this issue, contemporary ap- proaches [8, 14, 16, 21, 25, 27, 35, 36] propose pruning re- dundant tokens. They trade acceptable performance degra- dation for a more cost-effective model. Knowledge distil- lation [11] and other techniques can further mitigate the re- sulting performance drop. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2092 However, a steep drop in performance is inevitable as pruning tokens further increases because both essential sub- ject and auxiliary context information drop significantly, es- pecially when the number of reserved tokens is closely be- low 10. Aggressive token pruning could lead to incomplete subject and background context loss, causing the wrong pre- diction, as shown in Fig. 1. Specifically, the background tokens containing sod help recognize the input image as a lawn mower rather than a folding chair. Meanwhile, miss- ing subject tokens make the baseball indistinguishable from a rugby ball. To regain adequate information from pruned tokens, EViT [16] and Evo-ViT [35] propose aggregating pruned tokens as one, as shown in Fig. 2 (b). Still, they ne- glect the discrepancy among these tokens, leading to feature collapse and hindering more aggressive token pruning. Towards more aggressive pruning, we argue that infor- mation in pruned tokens deserves better treatment. We did a toy experiment to answer what accuracy token pruning could achieve if it applied the reversed pruning policy in the first pruned transformer block as Fig. 3 shows. Taking dy- namicViT [25] as a case study, the performance of reversed policy is enough to bring extra accuracy complementary to the original one (denoted by bonus accuracy). Moreover, this phenomenon would become more significant as prun- ing continues (red line in Fig. 3.). To conserve the information from the pruned tokens, we propose a Joint Token Pruning & Squeezing (TPS) module to accommodate more aggressive compression of ViTs. TPS module utilizes a feature dispatch mechanism that squeezes essential features from pruned tokens into re- served ones, as shown in Fig. 2 (c). Firstly, based on the scoring result, the TPS module divides input tokens into two complementary subsets: the reserved and pruned sets. Sec- ondly, instead of discarding or collapsing tokens from the pruned set into a single one, we employ a unidirectional nearest-neighbor matching algorithm to dispatch each of them independently to the associated reserved token dubbed as the host token. This design reduces information loss without sacrificing computational efficiency. Subsequently, we apply a similarity-based fusing way to squeeze the fea- tures of matched pruned tokens into corresponding host to- kens while the non-selected reserved tokens remain identi- cal. This design reduces the context information loss while retaining a reasonable computation budget. We can easily achieve hardware-friendly constant shape inference when fixing the cardinality of the reserved token set. Furthermore, we introduce two flexible variants: the inter-block version dTPS and the intra-block version eTPS, which are essen- tially plug-and-play blocks for both vanilla ViTs and hybrid ViTs. We conduct extensive experiments on two datasets: Im- ageNet1K [4] and large fine-grained dataset iNaturalist 2019 [29] to prove our efficiency, flexibility, and robustness.Firstly, experiments under different token pruning settings demonstrate the superior performance of our TPS while op- erating more aggressive compression compared with token pruning [25] and token reorganization [16]; further compar- isons with state-of-the-art transformers [8,13,20,28,31,36, 39, 40] show our promising efficiency. Secondly, we man- ifest the flexibility of our TPS by integrating it into popu- lar ViTs, including both vanilla ViTs and hybrid ViTs. Fi- nally, the evaluations under the random token selection pol- icy confirm the higher robustness of our TPS. Overall, our contributions are summarized as follows: • We propose the joint Token Pruning & Squeezing (TPS) and its two variants: dTPS and eTPS, to con- serve the information of discarded tokens and facilitate more aggressive compression of vision transformers. • Extensive experiments demonstrate our higher perfor- mance compared with prior approaches. Especially while compressing GFLOPs of DeiT-small&tiny to 35%, our TPS outperforms baselines with accuracy improvements of 1%-6%. • Broadest experiments applying our method to vanilla ViTs and hybrid ViTs show our flexibility, while the analysis experiments prove that our TPS is more robust than token pruning and token reorganization.
van_Amsterdam_ASPnet_Action_Segmentation_With_Shared-Private_Representation_of_Multiple_Data_Sources_CVPR_2023	Abstract Most state-of-the-art methods for action segmentation are based on single input modalities or na ¨ıve fusion of mul- tiple data sources. However, effective fusion of comple- mentary information can potentially strengthen segmenta- tion models and make them more robust to sensor noise and more accurate with smaller training datasets. In order to improve multimodal representation learning for action segmentation, we propose to disentangle hidden features of a multi-stream segmentation model into modality-shared components, containing common information across data sources, and private components; we then use an attention bottleneck to capture long-range temporal dependencies in the data while preserving disentanglement in consecutive processing layers. Evaluation on 50salads, Breakfast and RARP45 datasets shows that our multimodal approach out- performs different data fusion baselines on both multiview and multimodal data sources, obtaining competitive or bet- ter results compared with the state-of-the-art. Our model is also more robust to additive sensor noise and can achieve performance on par with strong video baselines even with less training data.	1. Introduction Action segmentation is the task of predicting which ac- tion is occurring at each frame in untrimmed videos of com- plex and semantically structured human activities [18, 32]. While conventional methods for human action understand- ing focus on classification of short video clips [6, 27, 34], action segmentation models have to learn the semantics of This research was funded in part by the Wellcome/EPSRC Cen- tre for Interventional and Surgical Sciences (WEISS) [203145/Z/16/Z]; the Engineering and Physical Sciences Research Council (EPSRC) [EP/P012841/1]; and the Royal Academy of Engineering Chair in Emerg- ing Technologies Scheme. For the purpose of open access, the author has applied a CC BY public copyright licence to any author accepted manuscript version arising from this submission. Figure 1. Different paradigms for multi-source data fusion via (a) early fusion, (b) disentanglement of modality-shared and modality-specific representations (our model) and (c) late fusion; (d) Example from 50salads highlighting shared and private infor- mation that can be extracted from video and accelerometer data. While both modalities can detect the activation of relevant tools and common motion cues, RGB videos additionally capture funda- mental details about objects without acceleration sensors and their state ( e.g. chopped tomatoes), the overall spatial configuration and the localization of motion in the scene. Accelerometer signals, on the other hand, contain explicit and complementary informa- tion about 3D fine motion patterns of activated objects and their co-occurrence. In the presence of noise ( e.g. video occlusions) or other variability factors, some shared attributes could become part of the private space of the uncorrupted modality. all action classes as well as their temporal boundaries and contextual relations, which is challenging and requires the design of efficient strategies to capture long range temporal information and inter-action correlations. Recent methods for action segmentation input pre- computed low-dimensional visual features [6, 11] into dif- ferent long-range temporal processing units, such as tem- poral convolutions [11, 19], temporal self-attention [38, 45] or graph neural networks [46]. While these methods utilize This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2384 only video data, recent computer vision datasets have in- creasing availability of multiple synchronized data sources [9,24,32], some of which could be collected readily in real- case scenarios ( e.g. audio recordings [9], teleoperated robot kinematics [35]). Effective fusion of different data modali- ties or different ‘ views’ of the same modality (here we use the term ‘view’ to denote any different representation of the same data source) is not trivial and still a very active area of research, as potential advantages include higher recogni- tion performance, improved robustness to sensor noise and mitigating the need for large training datasets [3]. Action segmentation with multiple data sources has not been investigated as extensively as similar tasks like action classification. It has generally been addressed via na ¨ıve fusion strategies such as multimodal feature concatenation [5, 45] and prediction fusion [37], or limited to the fea- ture encoding stage [22]. However, sensor fusion can also benefit from long-range temporal modelling performed in later stages. Inspired by work on multimodal representa- tion learning [5, 20], we approach the problem implement- ing a multi-stream action segmentation model, one stream for each available data source, and disentangling their latent space into modality-shared versus modality-specific repre- sentations (Fig. 1b and 1d), aiming at learning more dis- criminative features and more robust action recognition. We assume that creating a shared feature space across data sources produces more abstract action representations and reduces over-fitting to modality-specific nuances and noise, while private features could retain useful complementary information for the downstream task. Instead of relying on adversarial mechanisms [41], autoencoders [5, 20] or generative approaches [20], we learn shared feature spaces with minimal model modification by minimizing Maxi- mum Mean Discrepancy (MMD) on partitions of the la- tent spaces to reduce the distance between their distribu- tions. In order to capture long-range temporal dependen- cies in the data while preserving feature disentanglement in consecutive processing layers, an attention bottleneck [26] is then integrated into the segmentation model and initial- ized with learned modality-shared features, allowing inde- pendent processing of all private features. We called the model ASPnet (Action Shared-Private network). Evaluation results of our model on three challeng- ing benchmark datasets show improvement over unimodal baselines and different fusion strategies using both multi- modal ( e.g. video and accelerometer) and multiview ( e.g. RGB and optical flow) inputs, leading to competitive or better results than the state-of-the-art. In addition, results suggest that ASPnet could generalize well to multiple data sources, improving its performance with growing number of inputs. Despite requiring synchronized recordings of multiple sensors, we demonstrated that our model is also more robust to additive input noise and can match the per-formance of strong video baselines with less data. In sum- mary, our contributions are the following: • We present ASPnet, a new multi-source activity recog- nition model to effectively exploit shared and com- plementary information contained in multiple data sources for robust action segmentation. ASPnet par- titions the latent representation of each modality and exploits a bottleneck mechanism to allow feature inter- action at multiple levels of abstraction while preserv- ing disentanglement. Additionally, modality fusion is influenced by long-range temporal dynamics captured at different scales. • We show the advantage of feature disentanglement to fuse not only multimodal data, but also multiple repre- sentations of the same modality. • We perform extensive ablation studies to evaluate ASPnet against strong baselines, different levels of noise and less training data. • We evaluate ASPnet on three challenging benchmark datasets and achieved competitive or better results than state-of-the-art models.
Wan_Learning_Neural_Duplex_Radiance_Fields_for_Real-Time_View_Synthesis_CVPR_2023	Abstract Neural radiance fields (NeRFs) enable novel-view synthe- sis with unprecedented visual quality. However, to render photorealistic images, NeRFs require hundreds of deep mul- tilayer perceptron (MLP) evaluations – for each pixel. This is prohibitively expensive and makes real-time rendering infeasible, even on powerful modern GPUs. In this paper, we propose a novel approach to distill and bake NeRFs into highly efficient mesh-based neural representations that are fully compatible with the massively parallel graphics render- ing pipeline. We represent scenes as neural radiance features encoded on a two-layer duplex mesh, which effectively over- comes the inherent inaccuracies in 3D surface reconstruc- tion by learning the aggregated radiance information from a reliable interval of ray-surface intersections. To exploit local geometric relationships of nearby pixels, we leverage screen- space convolutions instead of the MLPs used in NeRFs to achieve high-quality appearance. Finally, the performance of the whole framework is further boosted by a novel multi- view distillation optimization strategy. We demonstrate the effectiveness and superiority of our approach via extensive experiments on a range of standard datasets.	1. Introduction Reconstructing 3D scenes by a representation that can be ren- dered from unobserved viewpoints using only a few posed * Corresponding author.images has been a long-standing goal in the computer graph- ics and computer vision communities. Significant progress has recently been achieved by neural radiance fields (NeRFs) [26], which are capable of generating photorealistic novel views and modeling view-dependent effects such as specu- lar reflections. In particular, a radiance field is a volumetric function parameterized by MLPs that estimates density and emitted radiance at sampled 3D locations in a given direction. Differentiable volume rendering then allows the optimization of this function by minimizing the photometric discrepancy between the real observed color and the rendered color. Despite the unprecedented success and enormous practi- cal potential of NeRF and its various extensions [ 3,6,57], an inescapable problem is the high computational cost of rendering novel views. For instance, even using a powerful modern GPU, NeRF requires about 30 seconds to render a single image with 800 ×800 pixels, which prevents its use for interactive applications in virtual and augmented real- ity. On the other hand, the rapid development of NeRF has spawned abundant follow-up works that focus on optimiza- tion acceleration [ 18,28,54], generalization [ 5,47,53], and enabled different downstream tasks, including 3D styliza- tion [ 11,16,30], editing [ 24,50,55,56], or even perception [10,17]. Thus, a generalized method that can learn and ex- tract a real-time renderable representation given an arbitrary pretrained NeRF-based model, while maintaining high ren- dering quality, is highly desirable. The huge computational cost of rendering a NeRF rep- resentation mainly comes from two aspects: (1) For each individual pixel, NeRF requires sampling hundreds of loca- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8307 tions along the corresponding ray, querying density and then accumulating radiance using volume rendering; and (2) A large model size is required to represent the geometric details of complex scenes well, so the MLPs used in NeRF architec- ture are relatively deep and wide, which incurs significant computation for the evaluation of each point sample. Figure 2. NeRF surface.In this paper, we present an ap- proach that achieves high-fidelity real-time view synthesis by ad- dressing these issues. To avoid the dense sampling along each ray, one solution is to generate a geome- try proxy from a pretrained NeRF model, e.g., via marching cubes [25]. By exploiting the highly- optimized graphics pipeline, we could almost instantly obtain the sample location for each ray. How- ever, due to inaccuracies in the den- sity field around surfaces, the extracted mesh may not faith- fully represent the true underlying geometry and can contain artifacts, as shown in Figure 2. Dense local sampling could alleviate these errors to some degree, but cannot handle miss- ing geometries or occlusions. An alternative approach is to use rasterization for fast neural rendering [ 1,37,38,43] by directly baking neural features onto the surface of explicit the geometry or point cloud. This is usually accompanied by a deep CNN to translate rasterized features to colors and learn to resolve the existing errors, which is expensive to evaluate and can prevent real-time rendering. To significantly reduce the sampled numbers while effi- ciently handling the geometric errors, our firstkey technical innovation is to infer the final RGB color according to the radiance of duplex points along the ray. As illustrated in Figure 3, NeRFs represent a continuous density distribution along each ray. While it will generally be difficult to deter- mine the exact location of a surface along the ray, it will be easier to extract a reliable interval that contributes the most to the final prediction by using an under- and an overesti- mation of the geometry. Motivated by this idea, instead of selecting a specific location for appearance calculation, or performing expensive dense volume rendering in this inter- val, we represent the scene using learnable features at the two intersection points of a ray with the duplex geometry and use a neural network to learn the color from this ag- gregated duplex radiance information. Although only two sampled locations are considered, we found this proposed neural duplex radiance field to be robust in compensating for the errors of the geometry proxy even without a deep neural network, while effectively preserving the efficiency of rasterization-based approaches. The NeRF MLP has become the most standard architecture for most neural implicit repre- sentations [ 7,8,26,27,31,39]. Yet, with only a few points Figure 3. Motivation. The continuous density distribution of NeRF (curve above the ray) makes it difficult to identify the accurate location of a surface ( H) for appearance calculation. We thus seek to extract a reliable interval from the density field (between dashed lines), and learn the duplex radiance combinations to tolerate errors. considered along the ray, the MLP struggles to constrain the proposed neural duplex radiance field. Instead , we use a shal- low convolutional network, which can effectively capture the local geometric information of neighboring pixels, and leads to a considerably better rendering quality. Finally , we found that directly training the neural duplex radiance field from scratch will lead to noticeable artifacts. We therefore propose a multi-view distillation optimization strategy that enables us to effectively approximate the rendering quality of the original NeRF models. Remarkably, our method im- proves run-time performance by 10,000 times compared to the original NeRF while maintaining high-quality rendering.
Wang_Conflict-Based_Cross-View_Consistency_for_Semi-Supervised_Semantic_Segmentation_CVPR_2023	Abstract Semi-supervised semantic segmentation (SSS) has re- cently gained increasing research interest as it can re- duce the requirement for large-scale fully-annotated train- ing data. The current methods often suffer from the confir- mation bias from the pseudo-labelling process, which can be alleviated by the co-training framework. The current co-training-based SSS methods rely on hand-crafted per- turbations to prevent the different sub-nets from collaps- ing into each other, but these artificial perturbations can- not lead to the optimal solution. In this work, we propose a new conflict-based cross-view consistency (CCVC) method based on a two-branch co-training framework which aims at enforcing the two sub-nets to learn informative features from irrelevant views. In particular, we first propose a new cross-view consistency (CVC) strategy that encourages the two sub-nets to learn distinct features from the same input by introducing a feature discrepancy loss, while these dis- tinct features are expected to generate consistent prediction scores of the input. The CVC strategy helps to prevent the two sub-nets from stepping into the collapse. In addition, we further propose a conflict-based pseudo-labelling (CPL) method to guarantee the model will learn more useful infor- mation from conflicting predictions, which will lead to a sta- ble training process. We validate our new CCVC approach on the SSS benchmark datasets where our method achieves new state-of-the-art performance. Our code is available at https://github.com/xiaoyao3302/CCVC .	1. Introduction Among different vision tasks, semantic segmentation is a fundamental vision task that enables the network to under- stand the world [3,11,12,32,33]. In recent years, deep neu- *This work was done during an internship at Samsung Research China- Beijing. This work is supported by Australian Research Council (ARC DP200103223). †Corresponding authors. Figure 1. We compare the cosine similarity values between the features extracted by the two sub-nets of the traditional cross- consistency regularization (CCR) method and our CVC method. We also compare the prediction accuracies of the two methods, measured by mIoU. We show that our CVC method can prevent the two sub-nets from collapsing into each other and inferring the input from irrelevant views, while CCR cannot guarantee the in- ferred views are different. We show our new method can increase the perception of the model, which produces more reliable pre- dictions. The experiments are implemented on the original Pascal VOC dataset, under the 1/4 split partition with ResNet-101 as the backbone of the encoder. ral networks (DNNs) have shown great potential in seman- tic segmentation [18,31,57]. However, the success of DNNs is mainly due to the huge amount of annotated datasets. For the task of semantic segmentation, pixel-level annotations are often required, which means the annotators need to man- ually label up to hundreds of thousands of pixels per image. Therefore, it takes great effort to collect precisely labelled data for training DNNs [1, 27, 30]. Various semi-supervised learning (SSL) methods are proposed to tackle the problem, which aim at learning a network by using only a small set of pixel-wise precisely annotated data and a large set of unlabelled data for seman- tic segmentation [2, 34, 37, 53, 54, 58]. It is obvious that the information from the labelled data is very limited as the number of labelled data is far less than the number of un- labelled data. Therefore, it becomes a key issue to fully exploit the unlabelled data to assist the labelled data for the model training. One intuitive way to tackle this issue is pseudo- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19585 labelling [28,37,48]. However, SSL methods along this line may suffer from the so-called confirmation bias [48], which often leads to performance degradation due to the unsta- ble training process. Recently, consistency regularization- based SSL methods show promising performance [35, 38, 41, 46]. However, most of them rely on producing the pre- dictions of the weakly perturbed inputs to generate pseudo- labels, which are then used as the supervision to generate the predictions of the strongly perturbed inputs. Therefore, they still suffer from the confirmation bias issue. On the other hand, co-training is a powerful framework for SSL as it enables different sub-nets to infer the same instance from different views and transfer the knowledge learnt from one view to another through pseudo-labelling. Particularly, co-training relies on multi-view reference to increase the perception of the model, thus enhancing the reliability of the generated pseudo-labels [40]. Various semi-supervised semantic segmentation (SSS) approaches are based on co-training [10, 39]. The key point is how to prevent different sub-nets from collapsing into each other such that we can make correct predictions based on the in- put from different views. However, the hand-crafted pertur- bations used in most SSS methods cannot guarantee hetero- geneous features to be learned to effectively prevent sub- nets from stepping into a collapse. Facing the above-mentioned issue, in this work, we come up with a new conflict-based cross-view consistency (CCVC) strategy for SSS, which makes sure the two sub- nets in our model can learn for different features separately so that reliable predictions could be learned from two irrel- evant views for co-training, thus further enabling each sub- net to make reliable and meaningful predictions. In particu- lar, we first raise a cross-view consistency (CVC) approach with a discrepancy loss to minimize the similarity between the feature extracted by the two sub-nets to encourage them to extract different features, which prevents the two sub-nets from collapsing into each other. Then we employ the cross pseudo-labelling to transfer the knowledge learnt from one sub-net to another to improve the perception of the network to correctly reason the same input from different views, thus producing more reliable predictions. However, the discrepancy loss may introduce too strong a perturbation to the model that the feature extracted by the sub-nets may contain less meaningful information for the prediction, leading to inconsistent and unreliable predic- tions from the two sub-nets. This will incur the confirma- tion bias problem and thus harm the co-training of the sub- nets. To tackle this problem, we further propose a conflict- based pseudo-labelling (CPL) method, where we encourage the pseudo-labels generated by the conflicting predictions of each sub-net to have stronger supervision for the pre- diction of each other, to enforce the two sub-nets to make consistent predictions. Thereby, the useful features for theprediction could be preserved as well as the reliability of the predictions. In this way, hopefully, the influence of the confirmation bias can be reduced and the training process can be more stable. As shown in Fig. 1, we can see the similarity scores be- tween the features extracted from the two sub-nets of the cross-consistency regularization (CCR) model remain at a high level, indicating the reasoning views of CCR are kind of relevant. In contrast, our CVC method ensures the rea- soning views are sufficiently different and thus produces more reliable predictions. It should be mentioned that our CCVC method is com- patible with various existing data augmentation methods and it also benefits from an augmented training set with in- creased data diversity. The contributions of our work are summarized as below: • We introduce a cross-view consistency (CVC) strategy based on a co-training framework to make reliable pre- dictions, where we propose a feature discrepancy loss to enable the two-branch network to learn how to rea- son the input differently but make consistent predic- tions. • We further propose a new conflict-based pseudo- labelling (CPL) method based on our cross-view con- sistency strategy to enable the two sub-nets to learn more useful semantic information from conflicting predictions to produce reliable and consistent predic- tions, which leads to a more stable training process. • Our method achieves the state-of-the-art performance on the commonly used benchmark datasets, PASCAL VOC 2012 [16] and Cityscapes [13].
Vidit_Learning_Transformations_To_Reduce_the_Geometric_Shift_in_Object_Detection_CVPR_2023	Abstract The performance of modern object detectors drops when the test distribution differs from the training one. Most of the methods that address this focus on object appearance changes caused by, e.g., different illumination conditions, or gaps between synthetic and real images. Here, by con- trast, we tackle geometric shifts emerging from variations in the image capture process, or due to the constraints of the environment causing differences in the apparent geometry of the content itself. We introduce a self-training approach that learns a set of geometric transformations to minimize these shifts without leveraging any labeled data in the new domain, nor any information about the cameras. We evalu- ate our method on two different shifts, i.e., a camera’s field of view (FoV) change and a viewpoint change. Our results evidence that learning geometric transformations helps de- tectors to perform better in the target domains.	1. Introduction While modern object detectors [1, 2, 17, 23, 24] achieve impressive results, their performance decreases when the test data depart from the training distribution. This prob- lem arises in the presence of appearance variations due to, for example, differing illumination or weather conditions. Considering the difficulty and cost of acquiring annotated data in the test (i.e., target) domain, Unsupervised Domain Adaptation (UDA) has emerged as the standard strategy to address such scenarios [3, 4, 9, 26, 38]. In this context, much effort has been made to learn do- main invariant features, such that the source and target dis- tributions in this feature space are similar. This has led to great progress in situations where the appearance of the ob- jects changes drastically from one domain to the other, as in case of real-to-sketch adaptation (e.g., Pascal VOC [10] to Comics [15]), or weather adaptation (e.g., Cityscapes [6] to Foggy Cityscapes [27]). Nevertheless, such object ap- pearance changes are not the only sources of domain shifts. They can also have geometric origins. For example, as shown in Fig. 1, they can be due to a change in camera view- Figure 1. Geometric shifts. (Left) Due to a different FoV , the cars highlighted in green, undergo different distortions even though they appear in similar image regions. (Right) Different camera viewpoints (front facing vs downward facing) yield dif- ferent distortions and occlusion patterns for pedestrian detection. (Bottom) The distributions of pedestrian bounding box sizes in Cityscapes [6] and MOT [8] differ significantly as the pedestrians are usually far away or in the periphery in Cityscapes. The top im- ages are taken from Cityscapes [6], and the bottom-left and right ones from KITTI [12] and MOT [8], respectively. point or field-of-view (FoV), or a change of object scale due to different scene setups. In practice, such geometric shifts typically arise from a combination of various factors, in- cluding but not limited to the ones mentioned above. In this paper, we introduce a domain adaptation approach tackling such geometric shifts. To the best of our knowl- edge, the recent work of [13] constitutes the only attempt at considering such geometric distortions. However, it intro- duces a method solely dedicated to FoV variations, assum- ing that the target FoV is fixed and known. Here, we de- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17441 velop a more general framework able to cope with a much broader family of geometric shifts. To this end, we model geometric transformations as a combination of multiple homographies. We show both the- oretically and empirically that this representation is suffi- cient to encompass a broad variety of complex geometric transformations. We then design an aggregator block that can be incorporated to the detector to provide it with the capacity to tackle geometric shifts. We use this modified detector to generate pseudo labels for the target domain, which let us optimize the homographies so as to reduce the geometric shift. Our contributions can be summarized as follows. (i) We tackle the problem of general geometric shifts for ob- ject detection. (ii) We learn a set of homographies using unlabeled target data, which alleviates the geometric bias arising in source-only training. (iii) Our method does not require prior information about the target geometric distor- tions and generalizes to a broad class of geometric shifts. Our experiments demonstrate the benefits of our approach in several scenarios. In the presence of FoV shifts, our approach yields similar performance to the FoV-dedicated framework of [13] but without requiring any camera infor- mation. As such, it generalizes better to other FoVs. Fur- thermore, we show the generality of our method by using it to adapt to a new camera viewpoint in the context of pedestrian detection.Our implementation can be accessed at https://github.com/vidit09/geoshift.
Wu_Uncovering_the_Disentanglement_Capability_in_Text-to-Image_Diffusion_Models_CVPR_2023	Abstract Generative models have been widely studied in computer vision. Recently, diffusion models have drawn substantial attention due to the high quality of their generated im- ages. A key desired property of image generative models is the ability to disentangle different attributes, which should enable modiﬁcation towards a style without changing the semantic content, and the modiﬁcation parameters should generalize to different images. Previous studies have found that generative adversarial networks (GANs) are inherently endowed with such disentanglement capability, so they can perform disentangled image editing without re-training or ﬁne-tuning the network. In this work, we explore whether diffusion models are also inherently equipped with such a capability. Our ﬁnding is that for stable diffusion models, by partially changing the input text embedding from a neu- tral description ( e.g., “a photo of person”) to one with style (e.g., “a photo of person with smile”) while ﬁxing all the Gaussian random noises introduced during the denoising process, the generated images can be modiﬁed towards the target style without changing the semantic content. Basedon this ﬁnding, we further propose a simple, light-weight image editing algorithm where the mixing weights of the two text embeddings are optimized for style matching and con- tent preservation. This entire process only involves optimiz- ing over around 50 parameters and does not ﬁne-tune the diffusion model itself. Experiments show that the proposed method can modify a wide range of attributes, with the performance outperforming diffusion-model-based image- editing algorithms that require ﬁne-tuning. The optimized weights generalize well to different images. Our code is publicly available at https://github.com/UCSB- NLP-Chang/DiffusionDisentanglement .	1. Introduction Image generation has been a widely-studied research problem in computer vision, with many competitive gen- erative models proposed over the last decade, such as gen- erative adversarial networks (GANs) [ 5,10,18,30,32] and variational autoencoders (V AE) [ 39,57,59,60]. Recently, diffusion models [ 23,71–73], with their ability to gener- ate high-quality and high-resolution images in different do- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1900 Scenes Person 4GlobalStyles (children drawing, cyberpunk, anime), Building appearance Styles (renaissance, Egyptian mural, sketch, Pixar) (wooden, red brick), Weather & time (sunset, night, snowy) Appearance (young, tanned, male) Local Cherry blossom, rainbow, foothills Expressions (smiling, crying, angry) 8Small edits Cake toppings, remove people on the street Hats, hair colors, earrings Table 1. Summarization of explored attributes. 4shows successfully disentangled attributes and 8shows failure cases. Small edits on the image are harder to be disentangled when the target attribute correlates with other parts of the image. mains, have soon attracted wide research attention. One important research direction regarding image gener- ative models is the ability to disentangle different aspects of the generated images, such as semantic contents and styles, which is crucial for image editing and style transfer. A generative model with a good disentanglement capability should satisfy the following two desirable properties. First, it should permit separate modiﬁcation of one aspect without changing other aspects. As an example shown in Fig. 2, in text-to-image generation, when the text input changes from “a photo of person” to “a photo of person with smile”, the generative model should have the ability to modify just the expression of the person ( i.e.,from the top image to mid- dle image in Fig. 2) without changing the person’s iden- tity (the bottom image in Fig. 2). Second, the parameters learned from modifying one image should transfer well to other similar images. For example, the optimal parameters that can add smile to one person should also work for im- ages of different people with different genders and races. Previous studies have discovered that GANs are inher- ently endowed with a strong disentanglement capability. Speciﬁcally, it is found that there exist certain directions in the latent space separately controlling different attributes. Therefore, by identifying these directions, e.g., via prin- cipal component analysis [ 19], GAN can achieve effective disentanglement without any re-training or ﬁne-tuning. On the other hand, such an inherent disentanglement capabil- ity has yet to be found in diffusion models. Hence come our research questions: Do diffusion models also possess a disentanglement capability with the aforementioned nice properties? If so, how can we uncover it? In this paper, we seek to answer these research questions. Our ﬁnding is that for stable diffusion model [ 61], one of the diffusion models that can generate images based on an in- put text description, disentangled image modiﬁcations can be achieved by partial modiﬁcations in the text embedding space. In particular, if we ﬁx the standard Gaussian noises introduced in the denoising process, and partially change the input text embedding from a neutral description ( e.g., “a photo of person”) to one with style ( e.g., “a photo of person with smile”), the generated image will also shift towards the target style without changing the semantic content. Based on this ﬁnding, we further propose a simple, light-weight al-gorithm, where we optimize the mixing weights of the two text embeddings under two objectives, a perceptual loss for content preservation and a CLIP-based style matching loss. The entire process only involves optimizing over around 50 parameters and does not ﬁne-tune the diffusion model. Our experiments show that the inherent disentanglement capability in stable diffusion model can already disentan- gle a wide range of concepts and attributes, ranging from global styles such as painting styles to local styles like fa- cial expressions, as shown in Table 1. As shown in Fig. 1, by learning the optimal mixing weights of the two de- scriptions, stable diffusion models can generate convinc- ing image pairs that only modify the target attribute, and the optimal weights can generalize well to different im- ages. The experiment results also show that our proposed image editing algorithm, without ﬁne-tuning the diffusion model, can match or outperform the more sophisticated diffusion-model-based image-editing baselines that require ﬁne-tuning. The ﬁndings of this paper can shed some light on how diffusion models work and how they can be applied to image editing tasks.
Wang_RIFormer_Keep_Your_Vision_Backbone_Effective_but_Removing_Token_Mixer_CVPR_2023	Abstract This paper studies how to keep a vision backbone ef- fective while removing token mixers in its basic building blocks. Token mixers, as self-attention for vision transform- ers (ViTs), are intended to perform information communica- tion between different spatial tokens but suffer from consid- erable computational cost and latency. However, directly removing them will lead to an incomplete model structure prior, and thus brings a significant accuracy drop. To this end, we first develop an R epIdentityFormer base on the re- parameterizing idea, to study the token mixer free model architecture. And we then explore the improved learn- ing paradigm to break the limitation of simple token mixer free backbone, and summarize the empirical practice into 5 guidelines. Equipped with the proposed optimization strat- egy, we are able to build an extremely simple vision back- bone with encouraging performance, while enjoying the high efficiency during inference. Extensive experiments and ablative analysis also demonstrate that the inductive bias of network architecture, can be incorporated into simple net- work structure with appropriate optimization strategy. We hope this work can serve as a starting point for the explo- ration of optimization-driven efficient network design.	1. Introduction The monumental advance in computer vision in the past few years was partly brought by the revolution of vision backbones, including convolutional neural networks (Con- vNets) [13,17,25,30] and vision transformers (ViTs) [14, 38]. Both of them have particular modules in their basic building blocks that aggregate information between differ- ent spatial locations, which are called token mixer [46], such as self-attention for ViTs. Although the effective- ∗Corresponding author. Token Mixer (Attention)LN LNChannel MLP Input Embedding ImageIdentity MappingLN LNChannel MLP Input Embedding ImageRemove token mixer 81.57ms232.42ms1666.01ms1725.61ms1874.02ms3070.33ms3096.26ms 46.3% Latency (a) Latency analysis of ViT-B (b) Remove token mixer with heavy latencyFigure 1. Latency analysis of different components in ViT- Base [ 14]. (a) For token mixer (self-attention), the latency oc- cupies about 46.3% of the backbone. (b) Our motivation was to remove the token mixer while striving to keep the performance. ness of token mixer has been demonstrated on many vision tasks [ 5,6,24,45], its computational complexity typically takes up a significant portion of the network. In practice, heavy token mixers make the vision backbone limited espe- cially on the edge-side devices due to the issue of speed and computation cost. There have been several attempts in the literature to in- vestigate efficient token mixers for slimming vision back- bones [ 29,31,46]. Although those works have already achieve competitive performance with light-weight design, they do retain the token mixers, which brings non-negligible increase in latency, as illustrated in Fig. 1. The recent work [ 47] finds that removing the token mixer is possible but leads to performance degeneration. Those explorations in efficient token mixers inspire us to think that can we keep the vision backbone effective but removing the token mixer? The resulting token mixer free vision backbone is expected to be efficient and effective for the realistic application. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14443 In this work, we first review the current model architec- tures and learning paradigms. Most of the previous works concentrate on the improvement of the architecture while adopting the conventional supervised learning to optimize the model from scratch. Differently, we propose to adopt the simplified model architecture, and explore the learning paradigm design to fully exploit the potential of the simple model. We aim to simultaneously maintain the efficiency and efficacy of token mixer free vision backbone (namely IdentityFormer, in Fig. 1-(b)). To this end, we investigate the simple and yet effective learning strategy, knowledge distillation (KD) [18] thoroughly in the following sections. Our main idea is distilling the knowledge from powerful teacher model (with token mixer) to the student model (to- ken mixer free). We instantiate the re-parameterizing idea to enlarge the modeling capacity of student network but retain its efficiency, as shown in Fig. 2. Specifically, the simple affine transformation is introduced into student model, to re- place the token mixer for training. The parameters of affine transformation can be merged into LayerNorm [ 2] during inference, which makes the student token mixer free finally. We empirically summarize the our learning strategy as the following guidelines, hope to shed light on how to learn the extremely simple model. Concretely, 1) soft distillation without using ground-truth labels is more effective; 2) using affine transformation without distillation is difficult to tailor the performance degeneration; 3) the proposed block-wise knowledge distillation, called module imitation , helps lever- aging the modeling capacity of affine operator; 4) teacher with large receptive field is beneficial to improve receptive field limited student; 5) loading the pre-trained weight of teacher model (except the token mixer) into student improve the convergence and performance. Based on the above guidelines, we finally obtain a to- ken mixer free vision model with competitive performance enjoying the high efficiency, dubbed as RepIdentityFormer (RIFormer) . RIFormer shares nearly the same macro and micro design as MetaFormer [ 46], but safely removing all token mixers. The quantitative results show that our net- works outperform many prevailing backbones with faster inference speed on ImageNet-1K [ 8]. And the ablative analyses on the feature distribution and Effective Recep- tive Fields (ERFs) also demonstrate that the inductive bias brought by an explicit token mixer, can be implicitly incor- porated into the simple network structure with appropriate optimization strategies. In summary, the main contributions of our work are as the following: • We propose to explore the vision backbone by develop- ing advanced learning paradigm for simple model archi- tecture , to satisfy the demand of realistic application. • We instantiate the re-parameterizing idea to build a to- ken mixer free vision model, RIFormer, which owns the improved modeling capacity for the inductive bias whileenjoying the efficiency during inference. • Our proposed practical guidelines of distillation strategy has been demonstrated effective in keeping the vision backbone competitive but removing the token mixer.
Wu_Incremental_3D_Semantic_Scene_Graph_Prediction_From_RGB_Sequences_CVPR_2023	Abstract 3D semantic scene graphs are a powerful holistic rep- resentation as they describe the individual objects and de- pict the relation between them. They are compact high-level graphs that enable many tasks requiring scene reasoning. In real-world settings, existing 3D estimation methods pro- duce robust predictions that mostly rely on dense inputs. In this work, we propose a real-time framework that incre- mentally builds a consistent 3D semantic scene graph of a scene given an RGB image sequence. Our method con- sists of a novel incremental entity estimation pipeline and a scene graph prediction network. The proposed pipeline simultaneously reconstructs a sparse point map and fuses entity estimation from the input images. The proposed net- work estimates 3D semantic scene graphs with iterative message passing using multi-view and geometric features extracted from the scene entities. Extensive experiments on the 3RScan dataset show the effectiveness of the pro- posed method in this challenging task, outperforming state- of-the-art approaches. Our implementation is available at https://shunchengwu.github.io/MonoSSG .	1. Introduction Scene understanding is a cornerstone in many computer vision applications requiring perception, interaction, and manipulation, such as robotics, AR/VR and autonomous systems [17, 54–56]. Semantic Scene Graphs (SSGs) go beyond recognizing individual entities (objects and stuff) by reasoning about the relationships among them [61, 66]. They also proved to be a valuable representation for com- plex scene understanding tasks, such as image caption- ing [26, 67], generation [13, 24], scene manipulation [10, 11], task planning [27], and surgical procedure estima- tion [42, 43]. Given the benefits of such representations, scene graph estimation received increasing attention in the computer vision community. While earlier methods mainly estimate SSGs from im- ages [18, 19, 33, 66, 72], recent approaches have also in- vestigated estimating them from 3D data. Compared to 2D scene graphs, which describe a single image, 3D scene input: RGB+Pose output: 3D boxes of instances (obj.+struc.) Sparse Entity Fusion Per-frame Entity estimation temporal label consistency global label merging Mapping Instance-level Segmentation 3DBBox extraction Labeling InSeg[3]++? ORBSLAM [1] ApproxMVBB [2] weighted inheritance Local graph: 3Dbbox. covis graph (green). neighbor braph (yellow) Global graph: 3Dbbox. covis graph (green). neighbor braph (yellow) Floor Sofaclose by Table Cushion close by Table Cushion stanting onstanding on supported byRGB sequence Sparse Reconstruction and Scene Graph Prediction at frame: 7 frame: 93 frame: 7 Incremental entity estimation and 3D scene graph estimation Figure 1. We propose a real-time 3D semantic scene graph estima- tion method that relies on an abstract understanding of a scene ge- ometry built with RGB input. Our method estimates scene graphs incrementally by continuously estimating scene graphs and fusing local predictions into a global 3D scene graph. graphs depict the entire 3D scenes, enabling applications requiring a holistic understanding of the whole scene, such as path planning [47], camera localization, and loop clo- sure detection [23]. However, existing 3D methods either require dense 3D geometry of the scenes to estimate 3D scene graphs [1, 23, 61, 64], which limits the use case since dense geometry is not always available, or constraints the scene graph estimation at the image-level [15,27,66], which tend to fail inferring relationships among objects beyond the individual viewpoints. A method that estimates 3D scene graphs relies on sparse scene geometry and reasoning about relationships globally has not been explored yet. In this work, we propose a real-time framework that in- crementally estimates a global 3D SSG of a scene simply requiring an RGB sequence as input. The process is illus- trated in Fig. 1. Our method simultaneously reconstructs a segmented point cloud while estimating the SSGs of the current map. The estimations are bound to the point map, which allows us to fuse them into a consistent global scene This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 5064 graph. The segmented map is constructed by fusing en- tity estimation from images to the points estimated from a sparse Simultaneous Localization and Mapping (SLAM) method [3]. Our network takes the entities and other proper- ties extracted from the segmented map to estimate 3D scene graphs. Fusing entities across frames is non-trivial. Exist- ing methods often rely on dense inputs [38,58] and struggle with sparse inputs since the points are not uniformly dis- tributed. Estimating scene graphs with sparse input points is also challenging. Sparse and ambiguous geometry ren- ders the node representations unreliable. On the other hand, directly estimating scene graphs from 2D images ignores the relationship beyond visible viewpoints. We aim to over- come the aforementioned issues by proposing two novel approaches. First, we propose a confidence-based fusion scheme which is robust to variations in the point distribu- tion. Second, we present a scene graph prediction network that mainly relies on multi-view images as the node feature representation. Our approach overcomes the need for exact 3D geometry and is able to estimate relationships without view constraints. In addition, our network is flexible and generalizable as it works not only with sparse inputs but also with dense geometry. We comprehensively evaluate our method on the 3D SSG estimation task from the public 3RScan dataset [60]. We ex- periment and compare with three input types, as well as 2D and 3D approaches. Moreover, we provide a detailed abla- tion study on the proposed network. The results show that our method outperforms all existing approaches by a signif- icant margin. The main contributions of this work can be summarized as follows: (1) We propose the first incremen- tal 3D scene graph prediction method using only RGB im- ages. (2) We introduce an entity label association method that works on sparse point maps. (3) We propose a novel network architecture that generalizes with different input types and outperforms all existing methods.
Weber_Power_Bundle_Adjustment_for_Large-Scale_3D_Reconstruction_CVPR_2023	Abstract We introduce Power Bundle Adjustment as an expansion type algorithm for solving large-scale bundle adjustment problems. It is based on the power series expansion of the inverse Schur complement and constitutes a new family of solvers that we call inverse expansion methods. We theo- retically justify the use of power series and we prove the convergence of our approach. Using the real-world BAL dataset we show that the proposed solver challenges the state-of-the-art iterative methods and significantly acceler- ates the solution of the normal equation, even for reaching a very high accuracy. This easy-to-implement solver can also complement a recently presented distributed bundle adjust- ment framework. We demonstrate that employing the pro- posed Power Bundle Adjustment as a sub-problem solver significantly improves speed and accuracy of the distributed optimization.	1. Introduction Bundle adjustment (BA) is a classical computer vision problem that forms the core component of many 3D recon- struction and Structure from Motion (SfM) algorithms. It refers to the joint estimation of camera parameters and 3D landmark positions by minimization of a non-linear repro- jection error. The recent emergence of large-scale internet photo collections [1] raises the need for BA methods that are scalable with respect to both runtime and memory. And building accurate city-scale maps for applications such as augmented reality or autonomous driving brings current BA approaches to their limits. As the solution of the normal equation is the most time consuming step of BA, the Schur complement trick is usu- ally employed to form the reduced camera system (RCS). This linear system involves only the pose parameters and is significantly smaller. Its size can be reduced even more by using a QR factorization, deriving only a matrix square root of the RCS, and then solving an algebraically equivalent 1Technical University of Munich 2Munich Center for Machine Learning 3University of Oxford (a)Ladybug-1197 (b)Venice-1102 Figure 1. Power Bundle Adjustment (PoBA) is a novel solver for large-scale BA problems that is significantly faster and more memory-efficient than existing solvers. (a) Optimized 3D recon- struction of a Ladybug BAL problem with 1197 poses. PoBA- 32 (resp. PoBA- 64) is41% (resp. 36% ) faster than the best competing solver to reach a cost tolerance of 1%. (b) Optimized 3D recon- struction of a Venice BAL problem with 1102 poses. PoBA- 32 (resp. PoBA- 64) is71% (resp. 69% ) faster than the best compet- ing solver to reach a cost tolerance of 1%.PoBA is five times (resp. twice) less memory-consuming than√ BA(resp. Ceres). problem [4]. Both the RCS and its square root formulation are commonly solved by iterative methods such as the pop- ular preconditioned conjugate gradients algorithm for large- scale problems or by direct methods such as Cholesky fac- torization for small-scale problems. In the following, we will challenge these two families of solvers by relying on an iterative approximation of the inverse Schur complement. In particular, our contributions This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 281 are as follows: •We introduce Power Bundle Adjustment ( PoBA ) for ef- ficient large-scale BA. This new family of techniques that we call inverse expansion methods challenges the state-of-the-art methods which are built on iterative and direct solvers. •We link the bundle adjustment problem to the theory of power series and we provide theoretical proofs that justify this expansion and establish the convergence of our solver. •We perform extensive evaluation of the proposed ap- proach on the BAL dataset and compare to several state-of-the-art solvers. We highlight the benefits ofPoBA in terms of speed, accuracy, and memory- consumption. Figure 1 shows reconstructions for two out of the 97 evaluated BAL problems. •We incorporate our solver into a recently proposed distributed BA framework and show a significant im- provement in terms of speed and accuracy. •We release our solver as open source to facili- tate further research: https://github.com/ simonwebertum/poba
Wu_Discriminating_Known_From_Unknown_Objects_via_Structure-Enhanced_Recurrent_Variational_AutoEncoder_CVPR_2023	Abstract Discriminating known from unknown objects is an im- portant essential ability for human beings. To simulate this ability, a task of unsupervised out-of-distribution object de- tection (OOD-OD) is proposed to detect the objects that are never-seen-before during model training, which is ben- eficial for promoting the safe deployment of object detec- tors. Due to lacking unknown data for supervision, for this task, the main challenge lies in how to leverage the known in-distribution (ID) data to improve the detector’s discrim- ination ability. In this paper, we first propose a method of Structure-Enhanced Recurrent Variational AutoEncoder (SR-VAE), which mainly consists of two dedicated recurrent VAE branches. Specifically, to boost the performance of ob- ject localization, we explore utilizing the classical Lapla- cian of Gaussian (LoG) operator to enhance the structure information in the extracted low-level features. Meanwhile, we design a VAE branch that recurrently generates the aug- mentation of the classification features to strengthen the dis- crimination ability of the object classifier. Finally, to alle- viate the impact of lacking unknown data, another cycle- consistent conditional VAE branch is proposed to synthesize virtual OOD features that deviate from the distribution of ID features, which improves the capability of distinguishing OOD objects. In the experiments, our method is evaluated on OOD-OD, open-vocabulary detection, and incremental object detection. The significant performance gains over baselines show the superiorities of our method. The code will be released at https://github.com/AmingWu/SR-VAE.	1. Introduction Recent years have witnessed the rapid development of deep learning based object detection [5,12,34,36,41], which often follows a close-set assumption that the training and testing processes share the same category space. How- ever, the practical scenario is open and filled with unknown *Corresponding author. I D T r a i n i n g D a t a S R -V A E O O D T e s t i n g D a t a S R -V A E S y n t h e s i z e V i r t u a l O O D F e a t u r e s OO D I D ❌ ❌Figure 1. Discriminating known from unknown objects (as shown in green boxes) by synthesizing virtual OOD features. For OOD- OD, to alleviate the impact of lacking unknown data, we present an SR-V AE method to constrain the synthesized features (as shown in blue stars) to deviate from the distribution of ID features (as shown in orange). Meanwhile, we consider enhancing the discrimination of the classifier to reduce the risk of misclassifying the ID objects into the OOD category. Through these operations, the ability of distinguishing OOD objects could be improved significantly. objects, e.g., in Fig. 1, an autonomous vehicle may en- counter an unseen camel, presenting significant challenges for close-set assumption based detectors. To promote the safe application of detectors, a task of unsupervised out-of- distribution object detection (OOD-OD) [7] is recently pro- posed, which aims to detect the objects never-seen-before during training without accessing any auxiliary data. Towards unsupervised OOD-OD, since there is no aux- iliary data available for supervision, leveraging the known in-distribution (ID) data to enhance the detector’s discrimi- nation ability becomes the critical challenge. One feasible solution is to synthesize a series of virtual OOD features [7, 35] based on the ID data, which is beneficial for pro- moting the object detector to learn a clear decision bound- ary between ID and OOD objects. To this end, the work [7] attempts to synthesize virtual features from the low- likelihood region of the estimated class-conditional distri- bution. However, this method requires a large number of objects for each category to estimate the distribution, limit- ing its application to the case of few samples. As shown in Fig. 1, in this paper, we consider improv- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23956 ing the performance of OOD object detection from two per- spectives: One is to strengthen the discrimination capabil- ity of the object classifier for the known ID objects, which is conducive to reduce the risk of misclassifying the ID ob- jects into the OOD category. Another is to synthesize the virtual OOD features that significantly deviate from the dis- tribution of the ID features, which is instrumental in boost- ing the performance of distinguishing OOD objects from ID objects. To attain these two goals, we explore exploiting Variational AutoEncoder (V AE) [15, 20] to separately gen- erate the augmented ID features and virtual OOD features. Specifically, an approach of Structure-Enhanced Recur- rent Variational AutoEncoder (SR-V AE) is proposed, which mainly consists of two dedicated recurrent V AE branches. In general, an object detector should first localize objects. Then, an object classifier is used to discriminate these ob- jects [12, 36]. To improve the localization performance, en- hancing the object-related information in the extracted fea- tures is meaningful. To this end, after extracting the low- level features of an input image, we present to utilize the classical Laplacian of Gaussian (LoG) operator [23] to ob- tain structure-related information, which is used to fuse into the existing features to strengthen the localization ability. Next, in order to enhance the discrimination ability, in- spired by Invariant Risk Minimization [1], we explore con- structing a set of diverse environments to intensify the vari- ance of the input classification features. Concretely, a V AE module [17, 20, 43] is exploited to recurrently output mul- tiple augmented features of the classification features, i.e., the current output is taken as the input of the next iteration. Since the input of each iteration is different, by means of the variation operation, the diversity of the output features could be enlarged. Then, the discrimination ability is im- proved by minimizing the prediction discrepancy between the augmented features and the classification features. Finally, to alleviate the impact of lacking unknown data, we present a cycle-consistent conditional V AE [40] to syn- thesize virtual OOD features in the absence of paired super- vision samples. Concretely, to ameliorate the synthesized features to deviate from the distribution of ID features, we first insert label information in the latent space to force a de- terministic constrained representation. Meanwhile, by max- imizing the discrepancy between the synthesized features and the input features, the synthesized features could be facilitated to contain plentiful OOD-relevant information, which enhances the ability of distinguishing OOD objects. In the experiments, our method is separately evaluated on three different tasks. Extensive experimental results on mul- tiple datasets demonstrate the superiorities of our method. The contributions are summarized as follows: (1) For unsupervised OOD-OD, we observe that using the classical LoG operator could effectively enhance object- related information in the extracted low-level features.(2) To reduce the risk of misclassifying ID objects into the OOD category, we design a dedicated recurrent V AE to generate diverse augmented features of the input classifica- tion features, which is beneficial for improving the discrim- ination ability of the object classifier. (3) To alleviate the impact of lacking unknown data for supervision, we present a cycle-consistent conditional V AE to synthesize virtual OOD features, which is instrumental in distinguishing OOD objects from ID objects. (4) In the experiments, our method is evaluated on OOD- OD [7], open-vocabulary detection [33, 49], and incremen- tal object detection [22, 39]. Particularly, for OpenImages dataset [24], compared with the baseline method [7], our method significantly reduces FPR95 by around 13.73% .
Wang_PlaneDepth_Self-Supervised_Depth_Estimation_via_Orthogonal_Planes_CVPR_2023	Abstract Multiple near frontal-parallel planes based depth repre- sentation demonstrated impressive results in self-supervised monocular depth estimation (MDE). Whereas, such a repre- sentation would cause the discontinuity of the ground as it is perpendicular to the frontal-parallel planes, which is detri- mental to the identification of drivable space in autonomous driving. In this paper, we propose the PlaneDepth, a novel orthogonal planes based presentation, including vertical planes and ground planes. PlaneDepth estimates the depth distribution using a Laplacian Mixture Model based on orthogonal planes for an input image. These planes are used to synthesize a reference view to provide the self- supervision signal. Further, we find that the widely used resizing and cropping data augmentation breaks the or- thogonality assumptions, leading to inferior plane predic- tions. We address this problem by explicitly constructing the resizing cropping transformation to rectify the prede- fined planes and predicted camera pose. Moreover, we propose an augmented self-distillation loss supervised with a bilateral occlusion mask to boost the robustness of or- thogonal planes representation for occlusions. Thanks to our orthogonal planes representation, we can extract the ground plane in an unsupervised manner, which is impor- tant for autonomous driving. Extensive experiments on the KITTI dataset demonstrate the effectiveness and effi- ciency of our method. The code is available at https: //github.com/svip-lab/PlaneDepth .	1. Introduction Monocular depth estimation (MDE) is an important task in computer vision and it has tremendous potential applica- tions, such as autonomous driving. However, the expensive data and labels acquisition process restricts the data scale in supervised MDE [3, 4, 9, 16, 26, 27]. Thus, researchers *Corresponding Author.turn to solve the data constraints in supervised MDE with the self-supervised MDE framework by leveraging videos or stereo image pairs. Most of the early works in self-supervised MDE leverage a regression module to estimate pixel-wise depth map [11, 12, 15, 29, 36, 40] and warp the reference view image to the target view based on the estimated depth. Then, a photo- metric consistency loss is used to guide the learning of the depth regression module. However, these methods usually encounter the local minimum issue because of the locality of bilinear interpolation on the reference view. To avoid this issue, rather than using simple depth regression, multiple frontal-parallel planes based depth representation is intro- duced where depth space is divided into a fixed number of frontal-parallel planes, and the depth network learns to clas- sify which predefined plane each pixel belongs to [13, 14]. It has been shown that such representation could produce much sharper depth on the edges of the object. However, they are insufficient to represent the ground because the ground plane is perpendicular to these predefined frontal- parallel planes. As shown in Fig. 1, such vertical depth planes only solution would lead to discontinuity on the ground, which is obviously detrimental to the identification of drivable space in autonomous driving. Further, photo- metric consistency loss is applied to the weighted compo- sition of each plane-warped image, which is sub-optimal as the combination of different weights may lead to exactly the same color image [30], resulting in ambiguous solutions for depth plane classification. Considering the ground is perpendicular to the frontal- parallel planes, in this paper, we propose to leverage orthog- onal planes to represent the scene where the ground planes favor the depth estimation in the ground region. Further, we propose to model the depth as a mixture of Laplace distri- butions of orthogonal planes [38], where each Laplacian is centered at one plane. We compute the photometric consis- tency loss independently on the color image warped by each plane, resulting in a more deterministic and less ambiguous optimization objective compared with the weighted compo- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21425 ③Ours ①PladeNet ① ② ③ ① ② ③ ①① ③ Input ① ② ③②②DepthHintFigure 1. Monocular Depth Estimation Results . The zoomed-in visualizations and bird-eye-view colored point cloud show that our method can predict continuous depth in the ground region while preserving the sharp edges of objects. Compared with PladeNet [13], our prediction has smoother depth in the ground region. Compared with DepthHint [40], our method addresses the occlusion problem, which can be seen in figure that our prediction have eliminated depth artifacts in the edges of street lights and cars. sition strategy mentioned above [13,14], consequently lead- ing to better depth estimation results, as shown in Fig. 1. Moreover, we can extract the ground plane in an unsuper- vised manner thanks to our orthogonal planes representa- tion. Resizing and cropping are widely used as a data augmen- tation strategy in the stereo training setting of MDE [13,14]. However, it would destroy the orthogonality of our prede- fined planes, leading to inferior plane distributions estima- tion. To remedy this issue, in this paper, we deeply ana- lyze the effects of resizing and cropping on the world co- ordinates system. We explicitly compute the resizing and cropping transformation and use it to rectify the predefined planes and the predicted camera rotation, which eases the learning of plane distributions. We further propose to use neural positional encoding (NPE) for the resizing and crop- ping parameter and incorporate it into our PlaneDepth net- work, which improves the robustness of network training. The self-distillation strategy is commonly used to solve occlusion problems [13, 14] and improve the depth predic- tion results [29]. Post-processing is widely used to improve the final prediction [11], which can be used naturally to gen- erate more accurate self-distillation labels. In this paper, we propose to combine post-processing with self-distillation by using a bilateral occlusion mask to generate more accurate supervision of network training, which improves both accu- racy and efficiency of our method. We summarize our contributions as follows: 1. We propose the PlaneDepth, a novel orthogonal plane- based monocular depth estimation network, which favours the representation of both vertical objects and ground. Such representation leads to a much smoother depth for ground regions and would facilitate the iden- tification of drivable regions. 2. The depth within the scene is modeled by a mixture of Laplacian distributions, and the depth classifica-tion problem is cast as the optimizing the mixture of Laplace distribution, which avoids the ambiguity in color expectation based depth estimation and leads to more stable depth estimation. 3. An orthogonality-preserved data augmentation strat- egy is proposed, which improves the robustness of net- work training. 4. We combine post-processing with self-distillation by our augmented self-distillation, which improves both efficiency and accuracy.
Xu_Abstract_Visual_Reasoning_An_Algebraic_Approach_for_Solving_Ravens_Progressive_CVPR_2023	Abstract Visual Reasoning: An Algebraic Approach for Solving Raven’s Progressive Matrices Jingyi Xu1∗Tushar Vaidya2◦∗Yufei Wu2◦*Saket Chandra1Zhangsheng Lai3◦Kai Fong Ernest Chong1† 1Singapore University of Technology and Design 2Nanyang Technological University3Singapore Polytechnic jingyi xu@mymail.sutd.edu.sg tushar.vaidya@ntu.edu.sg yufei002@e.ntu.edu.sg laizhangsheng@sp.edu.sg {saket chandra,ernest chong }@sutd.edu.sg Abstract We introduce algebraic machine reasoning, a new rea- soning framework that is well-suited for abstract reasoning. Effectively, algebraic machine reasoning reduces the diffi- cult process of novel problem-solving to routine algebraic computation. The fundamental algebraic objects of interest are the ideals of some suitably initialized polynomial ring. We shall explain how solving Raven’s Progressive Matri- ces (RPMs) can be realized as computational problems in algebra, which combine various well-known algebraic sub- routines that include: Computing the Gr ¨obner basis of an ideal, checking for ideal containment, etc. Crucially, the additional algebraic structure satisfied by ideals allows for more operations on ideals beyond set-theoretic operations. Our algebraic machine reasoning framework is not only able to select the correct answer from a given answer set, but also able to generate the correct answer with only the question matrix given. Experiments on the I-RAVEN dataset yield an overall 93.2%accuracy, which significantly out- performs the current state-of-the-art accuracy of 77.0%and exceeds human performance at 84.4%accuracy.	1. Introduction When we think of machine reasoning, nothing captures our imagination more than the possibility that machines would eventually surpass humans in intelligence tests and general reasoning tasks. Even for humans, to excel in IQ tests, such as the well-known Raven’s progressive matrices (RPMs) [5], is already a non-trivial feat. A typical RPM instance is composed of a question matrix and an answer set; see Fig. 1. A question matrix is a 3×3grid of panels *Equal contributions.†Corresponding author. ◦This work was done when the author was previously at SUTD. Code: https://github.com/Xu-Jingyi/AlgebraicMR Figure 1. An example of RPM instance from the I-RA VEN dataset. The correct answer is marked with a red box. that satisfy certain hidden rules, where the first 8 panels are filled with geometric entities, and the 9-th panel is “miss- ing”. The goal is to infer the correct answer for this last panel from among the 8 panels in the given answer set. The ability to solve RPMs is the quintessential display of what cognitive scientists call fluid intelligence. The word “fluid” alludes to the mental agility of discovering new re- lations and abstractions [28], especially for solving novel problems not encountered before. Thus, it is not surprising that abstract reasoning on novel problems is widely hailed as the hallmark of human intelligence [6]. Although there has been much recent progress in ma- chine reasoning [15, 17, 30–33, 37, 38, 46, 47], a common criticism [9, 25, 26] is that existing reasoning frameworks have focused on approaches involving extensive training, even when solving well-established reasoning tests such as RPMs. Perhaps most pertinently, as [9] argues, reasoning tasks such as RPMs should not need task-specific perfor- This work is supported by the National Research Foundation, Sin- gapore under its AI Singapore Program (AISG Award No: AISG-RP- 2019-015) and under its NRFF Program (NRFFAI1-2019-0005), and by Ministry of Education, Singapore, under its Tier 2 Research Fund (MOE- T2EP20221-0016). This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6715 Answer selection Answer generation Question matrix Concept matrixAnswer set Stage 1 Algebraic representation Stage 2 Algebraic machine reasoningCompute Concept matrix with answer 𝒊𝒊 Select answer Invariance modules “Inverse” invariance modules Generate answer Invariance patterns extracted from the first two rows (Perceptual attribute values are obtained from standard object detection models)Figure 2. An overview of our algebraic machine reasoning framework, organized into 2 stages. mance optimization. After all, if a machine optimizes per- formance by training on task-specific data, then that task cannot possibly be novel to the machine. To better emulate human reasoning, we propose what we call “algebraic machine reasoning”, a new reasoning frame- work that is well-suited for abstract reasoning. Our frame- work solves RPMs without needing to optimize for perfor- mance on task-specific data, analogous to how a gifted child solves RPMs without needing practice on RPMs. Our key starting point is to define concepts as ideals of some suitably initialized polynomial ring. These ideals are treated as the “actual objects of study” in algebraic machine reasoning, which do not require any numerical values to be assigned to them. We shall elucidate how the RPM task can be realized as a computational problem in algebra involving ideals. Our reasoning framework can be broadly divided into two stages: (1) algebraic representation, and (2) algebraic machine reasoning; see Fig. 2. In the first stage, we rep- resent RPM panels as ideals, based on perceptual attribute values extracted from object detection models. In the sec- ond stage, we propose 4 invariance modules to extract pat- terns from the RPM question matrix. To summarize, our main contributions are as follows: • We reduce “solving the RPM task” to “solving a computational problem in algebra”. Specifically, we present how the discovery of abstract patterns can be realized very concretely as algebraic computations known as primary decompositions of ideals. • In our algebraic machine reasoning framework, we in- troduce 4 invariance modules for extracting patterns that are meaningful to humans. • Our framework is not only able to select the correct an- swer from a given answer set, but also able to generate answers without needing any given answer set . • Experiments conducted on RA VEN and I-RA VEN datasets demonstrate that our reasoning framework significantly outperforms state-of-the-art methods.
Xie_Poly-PC_A_Polyhedral_Network_for_Multiple_Point_Cloud_Tasks_at_CVPR_2023	Abstract In this work, we show that it is feasible to perform multi- ple tasks concurrently on point cloud with a straightforward yet effective multi-task network. Our framework, Poly-PC, tackles the inherent obstacles (e.g., different model architec- tures caused by task bias and conflicting gradients caused by multiple dataset domains, etc.) of multi-task learning on point cloud. Specifically, we propose a residual set abstrac- tion (Res-SA) layer for efficient and effective scaling in both width and depth of the network, hence accommodating the needs of various tasks. We develop a weight-entanglement- based one-shot NAS technique to find optimal architec- tures for all tasks. Moreover, such technique entangles the weights of multiple tasks in each layer to offer task-shared parameters for efficient storage deployment while providing ancillary task-specific parameters for learning task-related features. Finally, to facilitate the training of Poly-PC, we introduce a task-prioritization-based gradient balance al- gorithm that leverages task prioritization to reconcile con- flicting gradients, ensuring high performance for all tasks. Benefiting from the suggested techniques, models optimized by Poly-PC collectively for all tasks keep fewer total FLOPs and parameters and outperform previous methods. We also demonstrate that Poly-PC allows incremental learning and evades catastrophic forgetting when tuned to a new task.	1. Introduction With the advances in deep learning, modern architec- tures offer tremendous improvements in 3D understand- ing [29, 36, 39, 54], e.g., point classification, segmentation, and detection, etc. Nevertheless, these networks are inef- ficient when handling numerous tasks since they are often intended to accomplish a single task. Even if parallel com- puting can address this issue, the memory footprints and storage costs grow linearly with the number of networks, *Corresponding author.rendering them unaffordable with constrained resources. Multitask learning (MTL) [3, 12, 13] offers a solution to this difficulty. In vision tasks, MTL models have been pre- dominantly proposed to simultaneously perform depth esti- mation, surface normal estimation, and semantic segmenta- tion on an input image [14, 19, 52]. Besides, the joint part- of-speech tagging, chunking, and named-entity recognition for Natural Language Processing (NLP) have also been in- vestigated [8, 9]. Since a substantial piece of the network (i.e., the backbone) is shared among tasks, an MTL model offers benefits in terms of complexity, inference time, and learning efficiency. However, training multiple tasks for point cloud poses two key challenges: 1) In contrast to common vision tasks, where a backbone that performs well on image classification can be directly ported to other tasks, the backbone for point cloud tasks must be carefully developed. Consequently, it is not feasible for all point cloud tasks to directly share a single backbone. 2) Instead of using a multi-task dataset as input, we seek to jointly perform multiple tasks on point cloud with mul- tiple dataset domains as input. Thus, in such a circum- stance, multi-task learning would result in considerable dis- parities in directions and magnitude of different task gradi- ents, a phenomenon known as task interference or negative transfer [34]. Meanwhile, task difficulty induced by multi- ple dataset domains is also different, so task prioritization should be considered to prevent placing undue emphasis on easier tasks when optimizing the multi-task network. To address the first challenge, we introduce residual set abstraction (Res-SA) layer, a scalable point feature learning module that can adapt to requirements for a variety of tasks in terms of width and depth of the network. Simultane- ously, when multiple tasks are presented to us, to reduce the manpower loss caused by manually designing the network, we seek to find the optimal architecture for each task us- ing neural network search (NAS). Thus, we construct differ- ent search spaces (neighbour points number, group radius, width, and depth, etc.) for multiple tasks on Res-SA. Then, inspired by AutoFormer [4], BigNAS [49] and slimmable This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1233 networks [48], we propose weight-entanglement-based one- shot NAS technique that entangles the weights of different tasks in the same layer, enabling different tasks to share parameters in their common parts for efficient storage de- ployment and offering task-specific parameters for learn- ing task-related features. Moreover, unlike previous works that share or monopolize all parameters in one layer for all tasks, such strategy allows different tasks to share a certain proportion of parameters in a single layer, achieving fine- grained parameter sharing. For negative transfer, previous methods narrow down the problem to two types of differences (i.e., gradient magni- tudes and directions) among task gradients, and propose several algorithms [5, 6, 12, 18, 35, 50] to homogenize the differences. However, these methods exclusively focus on one aspect of task gradient differences and also disregard the difficulty of various tasks. Intuitively, it is conceiv- able for easy tasks to dominate learning while harder ones stagnate during multi-task training. In response, we intro- duce a task-prioritization-based gradient balance algorithm that resolves negative transfer as a whole by homogeniz- ing both gradient magnitudes and directions across tasks via task prioritization. Specifically, we evaluate task prior- itization using the previous loss record. Then, we use task prioritization to homogenize task gradient magnitudes and directions in the current epoch, endowing difficult task with larger grad scale and enabling direction of final merged gra- dient vector closer to that of the difficult task. Since the task prioritization is dynamically adjusted, our algorithm avidly concentrates on the difficult task learning at each epoch and ensures that each task converges to the optimal solution. Over the well-trained Poly-PC, we undertake an evolu- tionary search with model size constraint to identify promis- ing architectures for different tasks. Experiments show that the searched models with weights inherited from the super- net outperform several baselines and are comparable with the state-of-the-arts trained individually for specific tasks. We also demonstrate that Poly-PC allows incremental learn- ing and evades catastrophic forgetting when generalizing to a new task. Thus, Poly-PC is parameter-efficient and can scale up more gracefully as the number of tasks increases. The key contributions can be summarized as follows: 1) We propose Poly-PC to perform multi-task learning on point cloud. To the best of our knowledge, Poly-PC is the first framework that takes multiple dataset domains as in- put for multi-task learning on point cloud. 2) We intro- duce Res-SA layer that meets the needs of different tasks in both the width and depth of the network. 3) We de- velop weight-entanglement-based one-shot NAS technique to find optimal architectures for different tasks as well as shared parameters inside each layer for efficient storage. 4) We propose task-prioritization-based gradient balance algo- rithm that resolves negative transfer as a whole to promotethe training of Poly-PC. 5) We demonstrate that Poly-PC allows incremental learning with fewer parameters.
Wang_EfficientSCI_Densely_Connected_Network_With_Space-Time_Factorization_for_Large-Scale_Video_CVPR_2023	Abstract Video snapshot compressive imaging (SCI) uses a two- dimensional detector to capture consecutive video frames during a single exposure time. Following this, an effi- cient reconstruction algorithm needs to be designed to re- construct the desired video frames. Although recent deep learning-based state-of-the-art (SOTA) reconstruction al- gorithms have achieved good results in most tasks, they still face the following challenges due to excessive model complexity and GPU memory limitations: 1) these models need high computational cost, and 2) they are usually un- able to reconstruct large-scale video frames at high com- pression ratios. To address these issues, we develop an ef- ficient network for video SCI by using dense connections and space-time factorization mechanism within a single residual block, dubbed EfficientSCI . The EfficientSCI net- work can well establish spatial-temporal correlation by us- ingconvolution in the spatial domain and Transformer in the temporal domain , respectively. We are the first time to show that an UHD color video with high compression ra- tio can be reconstructed from a snapshot 2D measurement using a single end-to-end deep learning model with PSNR above 32 dB. Extensive results on both simulation and real data show that our method significantly outperforms all pre- vious SOTA algorithms with better real-time performance. The code is at https://github.com/ucaswangls/ EfficientSCI.git .	1. Introduction Traditional high-speed camera imaging methods usually suffer from high hardware and storage transmission cost. Inspired by compressed sensing (CS) [5, 9], video snapshot compressive imaging (SCI) [45] provides an elegant solu- tion. As shown in Fig. 2, video SCI consists of a hardware encoder and a software decoder. In the encoder part, multi- ple raw video frames are modulated by different masks and *Equal Contribution, †Corresponding Author Testing time (s) PSNR (dB) Figure 1. Comparison of reconstruction quality (average PSNR in dB on 6 benchmark grayscale datasets) and testing time of several SOTA deep learning based algorithms. Our proposed EfficientSCI achieves higher reconstruction quality with fewer parameters and shorter testing time. then integrated by the camera to get a compressed measure- ment, giving low-speed cameras the ability to capture high- speed scenes. For the decoding part, the desired high-speed video is retrieved by the reconstruction algorithm using the captured measurement and masks. So far, many mature SCI imaging systems [14, 24, 31] have been built, but for the decoding part, there are still many challenges. In particular, although the model-based methods [21, 43, 44] have good flexibility and can recon- struct videos with different resolutions and compression rates, they require long reconstruction time and can only achieve poor reconstruction quality. In order to improve the reconstruction quality and running speed, PnP-FFDNet [46] and PnP-FastDVDnet [47] integrate the pre-trained denois- ing network into an iterative optimization algorithm. How- ever, they still need a long reconstruction time on large- scale datasets, e.g., PnP-FastDVDNet takes hours to recon- struct a UHD video from a single measurement. By contrast, deep learning based methods [28,30,35,40] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18477 Recovered Color VideoOriginal Scene Masks . . . Proposed EfficientSCI NetworkCameraBayer Filter Grayscale MeasurementBayer Measurement Recovered Grayscale Video● . . . t . . . t . . . t . . . tFigure 2. Schematic diagram of grayscale and color video SCI. have better real-time performance and higher reconstruction quality. For example, BIRNAT [8] uses bidirectional recur- rent neural network and generative adversarial method to surpass model-based method DeSCI [21] for the first time. MetaSCI [39] has made some explorations for the model to adapt to different masks, which reduces the model train- ing time. DUN-3DUnet [40] and ELP-Unfolding [42] com- bine iterative optimization ideas with deep learning mod- els to further improve reconstruction quality. However, due to the high model complexity and insufficient GPU mem- ory, most existing deep learning algorithms cannot train the models required for reconstructing HD or large-scale videos. RevSCI [7] uses a reversible mechanism [2] to re- duce the memory used for model training, and can recon- struct HD video with a compression rate up to 24, but the model training time increases exponentially. In addition, the current reconstruction algorithms generally use convolution to establish spatial-temporal correlation. Due to the local connection of convolution, long-term dependencies cannot be well established, and the model cannot reconstruct data with high compression rates. In summary, model-based methods usually require long reconstruction time and can only achieve poor reconstruc- tion quality. Learning-based methods have high model complexity but cannot be well applied to large-scale color video reconstruction. To address these challenges, we de- velop an efficient network for video SCI by using dense connections and space-time factorization mechanism . As shown in Fig. 1, our proposed method dramatically outper- forms all previous deep learning based reconstruction algo- rithms in terms of reconstruction quality and running speed with fewer parameters. Our main contributions can be sum- marized as follows: • An efficient end-to-end network, dubbed EfficientSCI, is proposed for reconstructing high quality video frames from a snapshot SCI measurement. • By building hierarchical dense connections within a single residual block, we devise a novel ResDNet block to effectively reduces model computational com- plexity but enhance the learning ability of the model.• Based on the space-time factorization mechanism, a Convolution and Trans former hybrid block (CFormer) is built, which can efficiently establish space-time cor- relation by using convolution in the spatial domain and Transformer in the temporal domain, respectively. • Experimental results on a large number of simu- lated and real datasets demonstrate that our proposed method achieves state-of-the-art (SOTA) results and better real-time performance.
Vendrow_JRDB-Pose_A_Large-Scale_Dataset_for_Multi-Person_Pose_Estimation_and_Tracking_CVPR_2023	Abstract Autonomous robotic systems operating in human envi- ronments must understand their surroundings to make ac- curate and safe decisions. In crowded human scenes with close-up human-robot interaction and robot navigation, a deep understanding of surrounding people requires reason- ing about human motion and body dynamics over time with human body pose estimation and tracking. However, ex- isting datasets captured from robot platforms either do not provide pose annotations or do not reﬂect the scene distri- bution of social robots. In this paper, we introduce JRDB- Pose, a large-scale dataset and benchmark for multi-person pose estimation and tracking. JRDB-Pose extends the ex- isting JRDB which includes videos captured from a social navigation robot in a university campus environment, con- taining challenging scenes with crowded indoor and out- door locations and a diverse range of scales and occlu- sion types. JRDB-Pose provides human pose annotations with per-keypoint occlusion labels and track IDs consistent across the scene and with existing annotations in JRDB. We conduct a thorough experimental study of state-of-the- art multi-person pose estimation and tracking methods on JRDB-Pose, showing that our dataset imposes new chal- lenges for the existing methods. JRDB-Pose is available at https://jrdb.erc.monash.edu/ .	1. Introduction Visual scene understanding of human environments is a difﬁcult and crucial task for autonomous driving, human- robot interaction, safe robotic navigation, and human action recognition. Although rough predictions of human location are sufﬁcient for some applications, a deep understanding of crowded human scenes and close-up human-robot inter- action requires reasoning about human motion and body dynamics with human body pose estimation and tracking. Developing an AI model to predict human body pose is *Equal contribution Figure 1. JRDB-Pose provides high frequency annotations of tracks and body joints in long scenes of crowded indoor and out- door locations featuring dynamic motion and occlusion. made more difﬁcult by the varied and highly imbalanced range of human motion found in daily living environments, including a variety of scales, occlusions, and overlapping humans, representing a long-tailed distribution of human poses which is difﬁcult for existing methods. Human pose estimation and tracking is an active research area with many new large-scale datasets [ 13,16,26,46] contributing to signiﬁcant recent progress; however, these datasets do not primarily target robotic perception tasks in social navigation environments, and thus rarely reﬂect spe- ciﬁc challenges found in human-robot interaction and robot navigation in crowded human environments, e.g. shopping malls, university campus, etc. JRDB [ 28] previously introduced a large-scale dataset and a benchmark for research in perception tasks related to robotics in human environments. The dataset was captured using a social manipulator robot with a multi-modal sensor suite including a stereo RGB 360° cylindrical video stream, 3D point clouds from two LiDAR sensors, audio and GPS positions. JRDB [ 28] additionally introduced annotations for 2D bounding boxes and 3D oriented cuboids. Recently, JRDB-Act [ 15] further introduced new annotations on the JRDB videos for individual actions, human social group This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4811 Dataset # Poses # Boxes Tracks Crowd ppF Occlusion Action Indoor + OutdoorRobot NavigationMulti- ModalMulti- Task MPII [ 3] 40K 1-17 X Penn Action [ 54] 160k 160k 1 XX COCO [ 26] 250k 500k 1-20 XX X KITTI [ 18] 80k XX X X X H3D [ 35] 460k XX X MOT20 [ 12] 1.65M XX X X TH¨OR [ 41] 2.5M XX X PoseTrack21 [ 13] 177k 429k XX 1-13 XX Waymo [ 46] 173K 9.9M XX unk XX X JRDB-Pose 636k 2.8M XX 1-36 XX X X XX JTA†[16] 10M X 0-60 X X MotSynth†[16] 40M X X 0-125 X X Table 1. Comparison of existing public datasets related to 2D pose estimation and tracking. For each dataset we report the numbers of poses, boxes, as well as the availability of tracking information, crowd data, people per frame (ppF), occlusion labels, action labels, scene type, and if the data comes from robot navigation in human environments. We mark if a dataset has data modalities besides RGB frames, and if it contains annotations for multi-task types. Note that JRDB-Pose is a multi-modal dataset captured from a social navigation robot, addressing different research challenges than many existing works.†Synthetic dataset. unk: Unknown. formation, and social activity of each social group. JRDB was collected from a robotic navigation platform in crowded human environments, diversely capturing both indoor and outdoor scenes. Additionally since the robot’s camera is located at person-level, and moves around, the data is not just collected from a far-off view but captures close-up scenes. For robotic systems to safely navigate dynamic human environments and perform collision risk prediction, they must be able to accurately track and forecast motion of people in their surroundings. Human motion is often fast and requires high frame rate data for accurate prediction and tracking, making high-frequency annotated human pose data crucial for the development and evaluation of robotic perception systems in human environments. Complex so- cial interactions add difﬁculty and similarly beneﬁt from high-frequency data. In crowded scenes with high levels of occlusions or overlap with other humans, tracking may be also difﬁcult. We introduce JRDB-Pose, a large-scale dataset captured from a mobile robot platform containing human pose and head box annotations. JRDB-Pose provides 600k pose an- notations and 600k head box annotations, each with an as- sociated tracking ID. JRDB-Pose includes a wide distri- bution of pose scales and occlusion levels, each with per- keypoint occlusion labels and consistent tracking IDs across periods of occlusion. The combination of JRDB-Pose with JRDB and JRDB-Act forms a valuable multi-modal dataset providing a comprehensive suite of annotations suited for robotic interaction and navigation tasks. Our contributions are: •We introduce JRDB-Pose, a large-scale pose estima- tion and tracking dataset providing pose annotations and head boxes with tracking IDs and per-keypoint oc- clusion labels. •In addition to adopting the popular metrics, we intro-duce new metrics, OSPA-Pose and OSPA(2)-Pose for pose estimation and tracking, respectively. •We conduct a comprehensive evaluation of state- of-the-art methods on JRDB-Pose and discuss the strengths and weaknesses of existing methods.
Wei_Enhancing_the_Self-Universality_for_Transferable_Targeted_Attacks_CVPR_2023	Abstract In this paper, we propose a novel transfer-based tar- geted attack method that optimizes the adversarial pertur- bations without any extra training efforts for auxiliary net- works on training data. Our new attack method is pro- posed based on the observation that highly universal ad- versarial perturbations tend to be more transferable for targeted attacks. Therefore, we propose to make the per- turbation to be agnostic to different local regions within one image, which we called as self-universality. Instead of optimizing the perturbations on different images, opti- mizing on different regions to achieve self-universality can get rid of using extra data. Specifically, we introduce a feature similarity loss that encourages the learned pertur- bations to be universal by maximizing the feature similar- ity between adversarial perturbed global images and ran- domly cropped local regions. With the feature similarity loss, our method makes the features from adversarial per- turbations to be more dominant than that of benign im- ages, hence improving targeted transferability. We name the proposed attack method as Self-Universality (SU) attack. Extensive experiments demonstrate that SU can achieve high success rates for transfer-based targeted attacks. On ImageNet-compatible dataset, SU yields an improvement of 12% compared with existing state-of-the-art methods. Code is available at https://github.com/zhipeng- wei/Self-Universality .	1. Introduction It has been demonstrated in recent works that adversar- ial examples have the properties of transferability, which means an adversarial example generated on one white- box model can be used to fool other black-box models [3, 15, 27, 30, 33]. The existence of transferability brings convenience to performing black-box attacks, hence raising security concerns for deploying deep models in real-world †Corresponding author.applications [14,22,28,35]. Consequently, considerable re- search attention has been spent on improving the transfer- ability of adversarial examples for both non-targeted and targeted attacks [4, 29, 36]. Compared to non-targeted attacks, transfer-based tar- geted attacks are inherently much more challenging since the goal is to fool deep models into predicting the specific target class. The major difficulty of transfer-based targeted attacks is caused by the fact that the gradient directions from a source image to a target class are usually different among different DNNs [16]. Hence, transfer-based attack methods designed for non-targeted attacks typically work poorly for targeted attacks. To increase the transferabil- ity, previous studies make efforts in aligning the feature of the generated adversarial example with the feature distribu- tions of the targeted class, which are learned from class- specific auxiliary networks [7, 8] or generative adversarial networks [18]. However, these works assume that the train- ing dataset is available and require extra training efforts for auxiliary networks, making it hard to apply in real-world scenarios. This paper investigates the problem of transfer-based tar- geted attacks. Specifically, we propose a new method that improves the transferability of adversarial examples in a more efficient way, i.e., without any training efforts for auxiliary networks to learn the feature distributions of the targeted class. Our method is proposed based on the ob- servation that more universal perturbations yield better at- tack success rates in targeted attacks. To this end, our goal is to enhance the universality of the generated adver- sarial perturbations, in order to improve its targeted trans- ferability. Note that existing universal adversarial pertur- bation (UAP) attacks [17] require optimizing the pertur- bations on an abundant of images to achieve universality, which is not applicable in our setting. To get rid of us- ing extra data and make transfer-based targeted attacks as convenient as non-targeted attacks, we propose to make the perturbation to be agnostic to different local regions within one image, which we called as self-universality. Then our method optimizes the self-universality of adversarial pertur- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12281 Shared Perturbations CNNs Local Input Random Cropping Resizing Data AugmentationGlobal Input Feature Similarity Loss Classification LossGlobal Feature Local Feature 𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎𝒎 𝒄𝒄𝒄𝒄𝒄𝒄𝜽𝜽𝜽𝜽Score Global PredictionTarget Class Local PredictionTotal Loss𝝀𝝀×Perturbation OptimizationFigure 1. Overview of the proposed SU attack. The random cropping is applied to the given benign image to generate the local image patch. After cropping, the local patch is resized to the shape of the benign image. Then both benign and local adversarial images with the shared perturbations are input to a surrogate white-box CNN model. Finally, the gradients obtained from the classification loss and the feature similarity loss are used to optimize perturbations. bations instead. To be specific, in addition to classification loss, our Self-Universality (SU) attack method introduces a feature similarity loss that maximizes the feature similarity between adversarial perturbed global images and randomly cropped local regions to achieve self-universality. In this way, our method makes the features from adversarial per- turbations to be more dominant than that of benign images, hence improving targeted transferability. Figure 1 gives an overview of the proposed Self- Universality (SU) attack. SU firstly applies random crop- ping on benign images to obtain local cropped patches. Then it resizes local patches to the same size with benign images. Consequently, global and local inputs with shared perturbations are input to the white-box model. Finally, ad- versarial perturbations are updated by minimizing the clas- sification loss ( e.g., Cross Entropy) between inputs and the target class and maximizing the feature similarity loss ( e.g., Cosine Similarity) of adversarial intermediate features be- tween local and global inputs. Benefiting from satisfying the prediction of the target class between global and lo- cal inputs and approximating adversarial intermediate fea- tures between the two, the proposed SU attack can generate perturbations with self-universality, thereby improving the cross-model targeted transferability. We briefly summarize our primary contributions as follows: • Through experiments, we find that highly universal ad- versarial perturbations tend to be more transferable for targeted attacks, which brings new insight into the de- sign of transfer-based targeted attack methods. • Based on the finding, we propose a novel Self- Universality (SU) attack method that enhances the uni- versality of adversarial perturbations for better targeted transferability without the requirement for extra data. • We conduct comprehensive experiments to demon-strate that the proposed SU attack can significantly im- prove the cross-model targeted transferability of adver- sarial images. Notably, SU can be easily combined with other existing methods.
Wang_Rethinking_the_Correlation_in_Few-Shot_Segmentation_A_Buoys_View_CVPR_2023	Abstract Few-shot segmentation (FSS) aims to segment novel ob- jects in a given query image with only a few annotated support images. However, most previous best-performing methods, whether prototypical learning methods or affinity learning methods, neglect to alleviate false matches caused by their own pixel-level correlation. In this work, we rethink how to mitigate the false matches from the perspective of representative reference features (referred to as buoys), and propose a novel adaptive buoys correlation (ABC) network to rectify direct pairwise pixel-level correlation, including a buoys mining module and an adaptive correlation module. The proposed ABC enjoys several merits. First, to learn the buoys well without any correspondence supervision, we customize the buoys mining module according to the three characteristics of representativeness, task awareness and re- silience. Second, the proposed adaptive correlation module is responsible for further endowing buoy-correlation-based pixel matching with an adaptive ability. Extensive experimen- tal results with two different backbones on two challenging benchmarks demonstrate that our ABC, as a general plu- gin, achieves consistent improvements over several leading methods on both 1-shot and 5-shot settings.	1. Introduction Semantic segmentation has achieved conspicuous achievements attributed to the recent advances in deep neural network [20]. However, its data-driven nature makes it heav- ily dependent on massive pixel-level training data, which is labor-intensive and time-consuming to collect. To imitate the human learning habits which can recognize new classes with only a glance, few-shot segmentation [25] (FSS) has attracted increasing interest in recent years, which aims at segmenting novel objects in the given query image with a few annotated support images. In previous literature, superior prototypical learning meth- ods [15, 28, 37] and affinity learning methods [11, 26, 30, 40] *Equal contribution †Corresponding author ② ② ③ False match ① ③ ②Buoys Distribution similarity between and : 0.9Distribution similarity between and : 0.2 ×Pixel similarity of and : 0.8 Pixel similarity of and : 0.9 (a) Pair -wise Correlation (b) Distribution based Correlation① Support foreground pixel Query foreground pixel Support background pixelPixel -level similarity Buoys correlationSupportQuery𝑝2𝑝1𝑝3 𝑝2𝑝1𝑝3Figure 1. The motivation of our proposed method. False matches tend to occur in the pixel-level correlation due to large intra-class variations. We introduce a series of representative features (buoys) as references and calculate the buoys-level correlation to suppress false matches. are almost all equipped with pixel-level correlation. In spe- cific, for prototypical learning methods, pixel-level correla- tion is implicitly endowed with the expectation to generate the foreground prior mask [28] for guiding the query pixel classification. For affinity learning methods, pixel-level cor- relation directly serves to aggregate support information and convey it to the query image [40]. Despite their promising results, these methods neglect the fact that there may exist cluttered background and inher- ent large intra-class variations between support and query images. In this case, directly employing pairwise pixel cor- relation may lead to considerable false matches. To make matters worse, the negative impact is inevitably amplified by inbuilt low-data regimes of FSS, leading to sub-optimal results. As shown in Fig. 1 (a), due to the significant pose difference of the object plane in the support-query image pair,p2in the query image located in the plane hatch is erro- neously closer to p3situated on the ground than counterpart p1in the support image. Therefore, it is highly desirable to rectify these false matches caused by the direct pairwise pixel-level correlation. In this paper, we aim to mitigate the false matches in previous FSS methods from the perspective of representative reference features (referred to as buoys ). Specifically, we design a novel Adaptive Buoys Correlation ( ABC ) network that can be applied as a generic plugin, including a buoy mining module and an adaptive correlation module to rec- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7183 tify direct pairwise pixel-level correlation for robust FSS. The main idea is, for each pixel from the support or query image, we can obtain the buoy-level correlation ( i.e., a like- lihood vector) by comparing this pixel with a set of buoys. In essence, the buoy-level correlation reflects the consensus among representative buoys with a broader receptive field, thus it encodes the relative semantic comparability of the buoys that can be relied upon. Intuitively, each pair of true pixel correlation ( e.g., the p1-p2pair in Fig. 1 (b)) derived from the query and support images should be not only vi- sually similar to each other ( i.e., high pairwise pixel-level correlation), but also similar to any other buoys ( i.e., simi- lar buoy-level correlation pair). Based on this correlation consistency in ABC, false matches caused by similar vision but dissimilar buoy correlations will be suppressed ( e.g., the point p3-p2pair in Fig. 1 (b)), ensuring that true pixel correlations enjoy higher weights to safely extract support information. However, it is non-trivial to learn the buoys well without any correspondence supervision for training. In the buoys mining module (BMM), we carefully design this module customized for the following three characteristics. (1) Rep- resentativeness. Intuitively, the buoys should have the ability to represent the diverse semantic clues from both support and query pixels with a broader semantic contrast descriptive. In other words, the matching between support-query pixels based on buoy-level correlation should preserve as much critical information as possible in the correspondence based on pixel-level correlation. In specific, we take advantage of Singular Value Decomposition (SVD) in pursuit of con- trollable information decay. Besides, a representation decay loss is devised to prevent the degradation of buoys. (2) Task awareness. Since tasks are randomly sampled during FSS training, each individual task consists of unique categories with large distribution differences. Therefore, to enable the buoys to perceive the current task and generalize to novel classes well, we employ the cross-aggregation mechanism to flexibly adjust buoys to meet the expectations of any tasks, even the tasks with unseen classes. (3) Resilience. Consider- ing the large gap between support and query images caused by large intra-class variations and cluttered background, it is necessary for buoys to bridge this gap and become more referable. In specific, we prepend the self-aggregation mech- anism to amend buoys by reconciling the intrinsic resilience between support and query images. Moreover, we observe that not all buoys are profitable when calculating pixel pair matching based on buoy-level correlation, and comprehensive consideration of the relation- ships between helpful buoys with potential intersections can assist in the final matching score. In the adaptive corre- lation module (ACM), we endow buoy-correlation-based pixel matching with an adaptive ability, which can flexibly assign less weight to irrelevant buoys conditioned to dif-ferent pixel pairs and focus on the structural similarity of related buoys. In specific, given the corresponding buoy pair for each pixel pair as the initial marginal distribution of the optimal transport(OT) algorithm, we can attain the opti- mal transport plan which can be regarded as the structural buoy contribution adaptive to the current pixel pair, and the corresponding OT distance is adopted for scoring matches. In this work, our contributions can be concluded as fol- lows: (1) We propose an Adaptive Buoys Correlation (ABC) network to rectify the widely used pairwise pixel correla- tion in FSS. To the best of our knowledge, this is the first work to mitigate the false matches in FSS methods, from the perspective of representative reference features (buoys). (2) We introduce two novel modules, namely Buoys Mining Module (BMM) and Adaptive Correlation Module (ACM), for representative buoys construction and adaptive matching respectively. They can cooperate well to achieve effective false match suppression. (3) Extensive experimental results with two different backbones on two challenging bench- marks demonstrate that our ABC, as a general plugin mod- ule, achieves consistent improvements over several leading methods on both 1-shot and 5-shot settings.
Wu_GANHead_Towards_Generative_Animatable_Neural_Head_Avatars_CVPR_2023	Abstract To bring digital avatars into people’s lives, it is highly demanded to efficiently generate complete, realistic, and animatable head avatars. This task is challenging, and it is difficult for existing methods to satisfy all the requirements at once. To achieve these goals, we propose GANHead (Generative Animatable Neural Head Avatar), a novel gen- erative head model that takes advantages of both the fine- grained control over the explicit expression parameters and the realistic rendering results of implicit representations. Specifically, GANHead represents coarse geometry, fine- gained details and texture via three networks in canonical space to obtain the ability to generate complete and realistic ∗Corresponding author Project page: https://wsj-sjtu.github.io/GANHead/head avatars. To achieve flexible animation, we define the deformation filed by standard linear blend skinning (LBS), with the learned continuous pose and expression bases and LBS weights. This allows the avatars to be directly ani- mated by FLAME [ 22] parameters and generalize well to unseen poses and expressions. Compared to state-of-the-art (SOTA) methods, GANHead achieves superior performance on head avatar generation and raw scan fitting.	1. Introduction How to efficiently generate photorealistic and animat- able head avatars without manual effort is an open prob- lem in computer vision and computer graphics, which has numerous applications in VR/AR, games, movies, and the metaverse. In these applications, it is desirable that the head This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 437 avatar models fulfill the following requirements: (1) Com- plete ,i.e., the 3D model can cover the entire head including the frontal face, the back of head, and the hair region; (2) Realistic , where the avatar is expected to display vivid tex- ture and detailed geometry; (3) Animatable ,i.e., the avatar is supposed to be fully riggable over poses and expressions, and can be controlled with low-dimensional parameters; (4) Generative model can be more flexibly applied to various downstream tasks, therefore, the head avatar model is pre- ferred to be generative rather than discriminative for large- scale content generation tasks. We investigate the research on neural head avatars and summarize previous works in Tab. 1. 3D morphable models (3DMMs) built from registered meshes have been widely employed to model head avatars. Principal component anal- ysis (PCA) is applied to shape and texture, and novel sub- jects can be generated by sampling the coefficients of PCA bases. However, registering real-world raw scans to a tem- plate mesh with fixed topology is non-trivial, and it is diffi- cult to define a fixed topology for complex regions like hair. As a result, most of these methods only model the facial region [ 3±6,13,15,23,39], while a few cover the full head without hair [ 2,11,22,33,36]. Moreover, the oversimplifi- cation of PCA makes the models lack of realism. In parallel with explicit meshes, implicit representations have been utilized to approximate complex surfaces. Some discriminative models [ 14,19,24,31,49] successfully model the complete head geometry with realistic texture. How- ever, these methods can only be applied to the reconstruc- tion task, incapable of generating new samples. Mean- while, 3D-aware GANs based on implicit representations [7,8,29,48] can generate multi-view-consistent frontal face images. Nevertheless, the heads are still incomplete. In addition, it is difficult to animate the neural head avatars generated by 3D-aware GANs. Recently, several implicit generative models [ 18,41,47,50] achieve realistic and ani- matable head avatars. However, these models either cannot generate complete head with satisfactory geometry [ 18,50], or can only be animated implicitly via the learned latent codes [ 47], which is inconvenient and limits the general- ization ability to unseen poses and expressions. It is natural to ask a question: can we build a model that can generate diverse realistic head avatars, and mean- while be compatible with the animation parameters of the common parametric face model (such as FLAME [ 22])? In this work, we propose a generative animatable neural head avatar model, namely, GANHead, that simultaneously fulfills these requirements. Specifically, GANHead repre- sents the 3D head implicitly with neural occupancy func- tion learned by MLPs, where coarse geometry, fine-gained details and texture are respectively modeled via three net- works. Supervised with unregistered ground truth 3D head scans, all these networks are defined in canonical space viaScheme Methods Complete Realistic Animatable Generative Explicit 3DMMs[2,3,15,40] ✗ ✗ ✓ ✓ 3D-aware GANs[7,8,29,48] ✗ ✓ ✗ ✓ Personalized Avatars[16,49] ✓ ✓ ✓ ✗ [14,31,44] ✗ ✓ △ ✗ Implicit Head Models[50] ✗ ✓ △ ✓ [18] ✗ ✓ ✓ ✓ [47] ✓ ✓ △ ✓ Ours ✓ ✓ ✓ ✓ Table 1. A summary of current head avatar methods. △denotes that the head avatar can only be animated implicitly via the learned latent codes, and cannot generalize well to unseen expressions. auto-decoder structures that are conditioned by shape, detail and color latent codes, respectively. This framework allows GANHead to achieve complete andrealistic generation re- sults, while yielding desirable generative capacity. The only remaining question is how to control the im- plicit representation with animation parameters? To answer this question, we extend the multi-subject forward skinning method designed for human bodies [ 9] to human faces, en- abling our framework to achieve flexible animation explic- itly controlled by FLAME [ 22] pose and expression param- eters. Inspired by IMAvatar [ 49], the deformation field in GANHead is defined by standard vertex based linear blend skinning (LBS) with the learned pose-dependent corrective bases, the linear blend skinning weights, and the learned expression bases to capture non-rigid deformations. In this way, GANHead can be learned from textured scans, and no registration or canonical shapes are needed. Once GANHead is trained, we can sample shape, de- tail and color latent codes to generate diverse textured head avatars, which can then be animated flexibly by FLAME parameters with nice geometry consistency and pose/expression generalization capability. We compare our method with the state-of-the-art (SOTA) complete head generative models, and demonstrate the superiority of our method. In summary, our main contributions are: • We propose a generative animatable head model that can generate complete head avatars with realistic tex- ture and detailed geometry. • The generated avatars can be directly animated by FLAME [ 22] parameters, robust to unseen poses and expressions. • The proposed model achieves promising results in head avatar generation and raw scan fitting compared with SOTA methods. 438
Worchel_Differentiable_Shadow_Mapping_for_Efficient_Inverse_Graphics_CVPR_2023	Abstract We show how shadows can be efficiently generated in differentiable rendering of triangle meshes. Our central ob- servation is that pre-filtered shadow mapping, a technique for approximating shadows based on rendering from the perspective of a light, can be combined with existing dif- ferentiable rasterizers to yield differentiable visibility infor- mation. We demonstrate at several inverse graphics prob- lems that differentiable shadow maps are orders of mag- nitude faster than differentiable light transport simulation with similar accuracy – while differentiable rasterization without shadows often fails to converge.	1. Introduction Differentiable renderers have become an essential tool for solving inverse problems in computer vision. They cur- rently come in two flavors: (1) forward rasterization us- inglocal shading models [9, 10, 39] and (2) path tracing and/or Monte Carlo methods for global light transport sim- ulation [22, 36, 53, 76]. While local methods are orders of magnitude faster, they lack effects of global light interaction such as shadows, caustics, or indirect illumination. Modern methods in real-time graphics can generate sur- prisingly realistic images by using efficient approximations of global effects. The single most important aspect for in- creasing the realism of local shading is the consideration of shadows (see Figure 1): for each pixel to be shaded, check if the path to a light source is unobstructed before evaluating a local shading model. Doing this accurately is costly and many approximate techniques have been developed. Our central observation is that one of the oldest and most widely used, shadow maps [71] (see Section 3), can be adapted to work in differentiable rendering frameworks. In Section 4 we explain how (certain approximations of) shadow mapping can be differentiated, exploiting ex- isting differentiable rasterizers. Our main idea is simi- lar to shadow maps: exploit efficient rasterization from the light’s point of view. For differentiable shadows this means: Existing differentiable rasterizers handle dis- Local Shading + Shadows Global Shading Local ShadingFigure 1. Adding shadows (middle) to local shading (left) is a significant step towards global light transport simulation (right). Rendering Numerical Deriva�ves Our Deriva�ves Figure 2. Complex scene (330k triangles) rendered in real time with our differentiable shadow mapping. It shows self-shadowing of objects, shadowing between objects, and colored surfaces. Fi- nite differences and our automatic derivatives w.r.t. movement of the light (note the non-zero derivatives at shadow boundaries). continuities of primary visibility along primitive borders; Primary Discon�nui�es Secondary Discon�nui�es we use this machinery to handle dis- continuities of secondary visibility along shadow borders. The resulting images contain shadows and are differentiable (see Figure 2). For many inverse graph- ics problems the level of realism they provide will suffice, while being gener- ated significantly faster than with global methods. This is important for machine learning tasks, where the renderings (and their derivatives w.r.t. scene parameters) are computed repeatedly. We provide de- tails of the implementation and how parameters affect opti- mization based on target images in Section 5. Given the importance of shadows for realistic image syn- thesis, it is unsurprising that many inverse problems heavily depend on them. We demonstrate the trade-off and possi- billities of differentiable shadow mappings in several appli- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 142 cations, ranging from pose estimation, over different types of geometry reconstruction, to interactive editing by manip- ulating shadows (Section 6). The main idea of this work, using existing differentiable rendering frameworks that readily resolve the problem of visibility discontinuities for dealing with discontinuities of secondary rays, may be useful for various other scenarios beyond shadows. We elaborate on this and other immediate consequences of our approach in Section 7.
Wang_BEV-LaneDet_An_Efficient_3D_Lane_Detection_Based_on_Virtual_Camera_CVPR_2023	Abstract 3D lane detection which plays a crucial role in vehicle routing, has recently been a rapidly developing topic in au- tonomous driving. Previous works struggle with practical- ity due to their complicated spatial transformations and in- ﬂexible representations of 3D lanes. Faced with the issues,our work proposes an efﬁcient and robust monocular 3D lane detection called BEV-LaneDet with three main contri-butions. First, we introduce the Virtual Camera that uniﬁes the in/extrinsic parameters of cameras mounted on differ-ent vehicles to guarantee the consistency of the spatial re- lationship among cameras. It can effectively promote thelearning procedure due to the uniﬁed visual space. We sec- ondly propose a simple but efﬁcient 3D lane representation called Key-Points Representation. This module is more suit- able to represent the complicated and diverse 3D lane struc-tures. At last, we present a light-weight and chip-friendlyspatial transformation module named Spatial Transforma- tion Pyramid to transform multiscale front-view features into BEV features. Experimental results demonstrate that our work outperforms the state-of-the-art approaches interms of F-Score, being 10.6% higher on the OpenLane dataset and 4.0% higher on the Apollo 3D synthetic dataset,with a speed of 185 FPS. Code is released at https: //github.com/gigo-team/bev_lane_det .	1. Introduction As one of the fundamental guarantees for autonomous driving, lane detection has recently received much attentionfrom researchers. Robust lane detection in real-time is oneof the foundations for advanced autonomous driving, which can provide substantial amounts of useful information for *Corresponding author This work was supported by the National Key Research and Devel- opment Project of New Generation Artiﬁcial Intelligence of China underGrant 2018AAA0102504.Autonomous Driving Systems (ADS), vehicle self-control, localization, and map construction. 3D Head Backbone STPBEV output Virtual CameraZ STP Figure 1. End-to-end framework illustrated. The original image in bottom-right is transformed to an input image by the Virtual Camera module. The input image is then encoded into front-view features by the backbone. Multiscale features from the backbone are put into the Spatial Transformation Pyramid (STP) to obtain BEV features. Given the BEV features, the 3D Head called Key- Points Representation generates BEV output and Z, the height of BEV lanes. At last, with the BEV output and Z, we can obtain 3D lanes. 2D lane detection methods have demonstrated remark- able performances [ 4,18,20,22]. Moreover, their outputs are usually projected to the ﬂat ground plane by Inverse Per-spective Transformation (IPM) with the camera in/extrinsic parameters, and then curve ﬁtting is performed to obtain the BEV lanes. However, the pipeline might cause other prob-lems in the actual driving process [ 1,20] for challenging situations like uphill and downhill. In order to overcome these problems, more recent meth- ods [ 3,6,8,9,14] have started to focus on the more com- plicated 3D lane perception domain. There are two signif- icant challenges in 3D lane detection: an efﬁcient spatialtransformation module to obtain BEV features and a robust This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1002 representation for 3D lane structures. The acquisition of BEV features is heavily dependent on camera in/extrinsic parameters, and previous methods have chosen to incor-porate camera in/extrinsic parameters into the network toobtain BEV features. Moreover, unlike obstacles on theroad, lane structures are slender and diverse. These meth-ods [ 3,8,9] carefully design the 3D anchor representation for lane structures with strong priors. However, they lacksufﬁcient ﬂexibility in some speciﬁc scenarios, as shown inFigure 5. Moreover, the anchor-free method [ 6] proposes a 3D lane representation based on the hypothesis that a linesegment in each predeﬁned tile is straight. This representa-tion is complicated and inaccurate. Towards the issues, we introduce BEV-LaneDet, an ef- ﬁcient and real-time pipeline that achieves 3D lane detec-tion from a single image, as shown in Figure 1. Different from incorporating camera in/extrinsic parameters into the network to get BEV features, we establish a Virtual Cam- era, which is applied to images directly. The module uniﬁes the in/extrinsic parameters of front-facing cameras in differ- ent vehicles by the homography method [ 2] based on BEV . This module guarantees the consistency of the spatial rela- tionship of front-facing cameras in different vehicles and re- duces variance in data distribution. Therefore, it can effec- tively promote the learning procedure due to the uniﬁed vi-sual space. We also propose a Key-Points Representation as our 3D lane representation. We demonstrate that it is a sim-ple but effective module to represent 3D lanes and is more expandable for complicated lane structures in some special scenarios. Moreover, the cost of computation and the chip’s friendliness are also crucial factors in autonomous driving.Therefore, a light-weight and easy-to-deploy spatial trans- formation module based on MLP is our preference. Mean-while, inspired by FPN [ 17], we present the Spatial Trans- formation Pyramid , which transforms multiscale front-view features to BEV and provides robust BEV features for 3D lane detection. In our experiments, we perform extensive studies to conﬁrm that our BEV-LaneDet signiﬁcantly out- performs the state-of-the-art PersFormer [ 3] in terms of F- Score, being 10.6% higher on the OpenLane real-world test set [ 3] and 4.0% higher on the Apollo simulation test set [ 9] with a speed of 185 FPS. In summary, our main contributions are three-fold: 1)Virtual Camera , a novel preprocessing module to unify the in/extrinsic parameters of cameras, ensuring data distri- bution consistency. 2)Key-Points Representation , a simple but effective representation of 3D lane structures. 3)Spa- tial Transformation Pyramid , a light-weight and easy-to- deploy architecture based on MLP to realize transformation from multiscale front-view features to BEV . Experimentsdemonstrate that our BEV-LaneDet achieves the state-of- the-art performance compared to other 3D lane detection algorithms.
Xiong_CAPE_Camera_View_Position_Embedding_for_Multi-View_3D_Object_Detection_CVPR_2023	Abstract In this paper, we address the problem of detecting 3D ob- jects from multi-view images. Current query-based methods rely on global 3D position embeddings (PE) to learn the ge- ometric correspondence between images and 3D space. We claim that directly interacting 2D image features with global 3D PE could increase the difficulty of learning view trans- formation due to the variation of camera extrinsics. Thus we propose a novel method based on CAmera view Position Embedding, called CAPE. We form the 3D position embed- dings under the local camera-view coordinate system instead of the global coordinate system, such that 3D position em- bedding is free of encoding camera extrinsic parameters. Furthermore, we extend our CAPE to temporal modeling by exploiting the object queries of previous frames and encod- ing the ego motion for boosting 3D object detection. CAPE achieves the state-of-the-art performance ( 61.0%NDS and 52.5%mAP) among all LiDAR-free methods on nuScenes dataset. Codes and models are available.1	1. Introduction 3D perception from multi-view cameras is a promising solution for autonomous driving due to its low cost and rich semantic knowledge. Given multiple sensors equipped on autonomous vehicles, how to perform end-to-end 3D percep- tion integrating all features into a unified space is of critical importance. In contrast to traditional perspective-view per- ception that relies on post-processing to fuse the predictions from each monocular view [44, 45] into the global 3D space, perception in the bird’s-eye-view (BEV) is straightforward and thus arises increasing attention due to its unified repre- sentation for 3D location and scale, and easy adaptation for downstream tasks such as motion planning. The camera-based BEV perception is to predict 3D ge- ometric outputs given the 2D features and thus the vital *Equal contribution.†Corresponding author. This work is done when Kaixin Xiong is an intern at Baidu Inc. 1Codes of Paddle3D and PyTorch Implementation. (a)PETRv2GlobalImagePESoftmax&MatMul+MatMul+ GlobalQueryPEVKQ (b)CAPE LocalImagePESoftmax&MatMulMatMul+MatMul LocalQueryPEVKQKQ ImageFeaturesDecoderEmbeddingFeatureGuidedPEFigure 1. Comparison of the network structure between PETRv2 and our proposed CAPE. (a) In PETRv2, position embedding of queries and keys are in the global system. (b) In CAPE, position embeddings of queries and keys are within the local system of each view. Bilateral cross-attention is adopted to compute attention weights in the local and global systems independently. challenge is to learn the view transformation relationship between 2D and 3D space. According to whether the explicit dense BEV representation is constructed, existing BEV ap- proaches could be divided into two categories: the explicit BEV representation methods and the implicit BEV repre- sentation methods. The former constructs an explicit BEV feature map by lifting the 2D perspective-view features to 3D space [12, 19, 34]. The latter mainly follow DETR-based [3] approaches in an end-to-end manner. Without projection or lift operation, those methods [24, 25, 52] implicitly encode the 3D global information into 3D position embedding (3D PE) to obtain 3D position-aware multi-view features, which is shown in Figure 1(a). Though learning the transformation from 2D images to 3D global space is straightforward, we reveal that the in- teraction in the global space for the query embeddings and 3D position-aware multi-view features hinders performance. The reasons are two-fold. For one thing, defining each cam- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21570 (a)ImagetoGlobal3D(b)ImagetoLocal3D (a)Imagetoglobal3D(b)Imagetolocal3D图五优化后优化前Figure 2. View transformation comparison. In previous methods, view transformation is learned from the image to the global 3D coordinate system directly. In our method, the view transformation is learned from image to local (camera) coordinate system. era coordinate system as the 3D local space, we find that the view transformation couples the 2D image-to-local trans- formation and the local-to-global transformation together. Thus the network is forced to differentiate variant camera extrinsics in the high-dimensional embedding space for 3D predictions in the global system, while the local-to-global relationship is a simple rigid transformation. For another, we believe the view-invariant transformation paradigm from 2D image to 3D local space is easier to learn, compared to directly transforming into 3D global space. For example, though two vehicles in two views have similar appearances in image features, the network is forced to learn different view transformations, as depicted in Figure 2 (a). To ease the difficulty in view transformation from 2D image to global space, we propose a simple yet effective approach based on local view position embedding, called CAPE, which performs 3D position embedding in the lo- cal system of each camera instead of the 3D global space. As depicted in Figure 2 (b), our approach learns the view transformation from 2D image to local 3D space, which eliminates the variances of view transformation caused by different camera extrinsics. Specially, as for key 3D PE, we transform camera frus- tum into 3D coordinates in the camera system using camera intrinsics only, then encoded by a simple MLP layer. As for query 3D PE, we convert the 3D reference points defined in the global space into the local camera system with camera extrinsics only, then encoded by a simple MLP layer. In- spired by [25, 29], we obtain the 3D PE with the guidance of image features and decoder embeddings, for keys and queries, respectively. Given that 3D PE is in the local space whereas the output queries are defined in the global coordi- nate system, we adopt the bilateral attention mechanism to avoid the mixture of embeddings in different representation spaces, as shown in Figure 1(b). We further extend CAPE to integrate multi-frame tempo- ral information to boost the 3D object detection performance,named CAPE-T. Different from previous methods that either warp the explicit BEV features using ego-motion [11, 19] or encode the ego-motion into the position embedding [25], we adopt separated sets of object queries for each frame and encode the ego-motion to fuse the queries. We summarize our key contributions as follows: •We propose a novel multi-view 3D detection method, called CAPE, based on camera-view position embed- ding, which eliminates the variances of view transfor- mation caused by different camera extrinsics. •We further generalize our CAPE to temporal modeling, by exploiting the object queries of previous frames and leveraging the ego-motion explicitly for boosting 3D object detection and velocity estimation. •Extensive experiments on the nuScenes dataset show the effectiveness of our proposed approach and we achieve the state-of-the-art among all LiDAR-free meth- ods on the challenging nuScenes benchmark.
Xue_IMP_Iterative_Matching_and_Pose_Estimation_With_Adaptive_Pooling_CVPR_2023	Abstract Previous methods solve feature matching and pose esti- mation using a two-stage process by ﬁrst ﬁnding matches and then estimating the pose. As they ignore the geomet- ric relationships between the two tasks, they focus on ei- ther improving the quality of matches or ﬁltering poten- tial outliers, leading to limited efﬁciency or accuracy. In contrast, we propose an iterative matching and pose es- timation framework (IMP) leveraging the geometric con- nections between the two tasks: a few good matches are enough for a roughly accurate pose estimation; a roughly accurate pose can be used to guide the matching by pro- viding geometric constraints. To this end, we implement a geometry-aware recurrent attention-based module which jointly outputs sparse matches and camera poses. Specif- ically, for each iteration, we ﬁrst implicitly embed geo- metric information into the module via a pose-consistency loss, allowing it to predict geometry-aware matches pro- gressively. Second, we introduce an efﬁcient IMP , called EIMP , to dynamically discard keypoints without potential matches, avoiding redundant updating and signiﬁcantly re- ducing the quadratic time complexity of attention computa- tion in transformers. Experiments on YFCC100m, Scannet, and Aachen Day-Night datasets demonstrate that the pro- posed method outperforms previous approaches in terms of accuracy and efﬁciency. Code is available at https: //github.com/feixue94/imp-release	1. Introduction Feature matching and relative pose estimation are two fundamental tasks in computer vision and especially impor- tant to visual localization and 3D reconstruction. Tradition- ally, the two tasks are performed in two stages separately by ﬁrst ﬁnding correspondences between keypoints ex- tracted from two images with nearest neighbor (NN) match- ing and then estimating the relative pose from predicted matches with robust estimators, e.g. RANSAC [6,7,20,32]. This pipeline has been the de-facto standard framework for decades [5]. However, due to repetitive textures/structures, 1𝟑𝟑𝟑𝟑/𝟓𝟓 .𝟑𝟑/𝟒𝟒.𝟔𝟔𝟏𝟏𝟏𝟏𝟑𝟑𝟒𝟒/𝟏𝟏𝟏𝟏𝟑𝟑𝟒𝟒 𝟑𝟑𝟔𝟔𝟑𝟑/𝟑𝟑𝟑𝟑𝟑𝟑𝟑𝟑𝟏𝟏𝟓𝟓/𝟑𝟑𝟓𝟓𝟑𝟑 𝟑𝟑𝟐𝟐𝟓𝟓/𝟑𝟑𝟓𝟓𝟑𝟑 23 4𝟗𝟗𝟑𝟑/𝟑𝟑 .𝟏𝟏/𝟏𝟏.𝟐𝟐𝟏𝟏𝟏𝟏𝟗𝟗/𝟏𝟏 .𝟏𝟏/𝟏𝟏.𝟐𝟐 𝟏𝟏𝟑𝟑𝟏𝟏/𝟏𝟏 .𝟔𝟔/𝟏𝟏.𝟑𝟑Figure 1. Process of iterative matching and pose esti- mation . For each image pair, we report the number in- liers/rotation/translation errors (top-left) and retained keypoints in left/right images (top-right) at iterations from 1 to 4. In the it- erative process, our method ﬁnds more inliers spanning almost the whole image, estimates increasingly precise pose and discards keypoints without true matches gradually. changing appearances and viewpoint variations, matches given by NN often contain a large number of outliers, lead- ing to poor pose accuracy [36, 37]. To mitigate this prob- lem, some works [8, 10, 16, 27, 40, 48, 49, 52] ﬁlter potential outliers of predicted matches with neural networks to im- prove the pose accuracy. Although they report better results, their performance is limited by the quality of initial matches and require extra time for ﬁltering at test time. Alterna- tively, advanced matchers such as SuperGlue [36] enhance the matching quality directly by using global information from all keypoints via transformers [45] with a ﬁxed num- ber ( e.g. 9) of iterations. These methods have obtained re- markable performance. Yet, their quadratic time complex- ity for the attention computation degrades the efﬁciency in real applications. Some following works [12,38,41] explore more efﬁcient variations, they run faster but are signiﬁcantly less accurate (see Table 1 and 3). In this paper, we aim to introduce an efﬁcient and accu- rate framework for iterative matching and pose estimation. Our approach is built upon the following observations: (1) a few well distributed matches ( e.g. 5) could give a roughly accurate pose ( e.g. essential matrix); (2) in turn, a roughly This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21317 accurate pose could provide strong geometric constraints (e.g. epipolar line) to ﬁnd more accurate matches at low cost; (3) the pose also reveals which keypoints have po- tential correspondences, preventing redundant operations. Based on the geometric connections of the two tasks, we propose an iterative matching and pose estimation frame- work (IMP), to perform matching and pose estimation it- eratively as opposed to in two separate stages. Speciﬁcally, we progressively augment descriptors with self and cross at- tention as [12,36,38,41], ﬁnd matches and estimate the rela- tive pose. As descriptors get gradually more discriminative, more correct matches can be found, leading to increasingly more precise pose, as shown in Fig. 1. However, due to the noise [26] and degeneration ( e.g. co-planar keypoints) [13], not all inliers could give a good pose [4, 18]. In addition to the classiﬁcation loss mainly used by prior methods [12,36], we apply a pose-consistency loss [49] to the matching pro- cess, enabling the model to ﬁnd matches which are not only accurate but also able to give a good pose. Moreover, in order to avoid redundant operations on un- informative keypoints, we employ a sampling strategy by combining the matching and attention scores of keypoints and the uncertainty of predicted poses to adaptively remove useful keypoints, as shown in Fig. 1. Compared with prior sampling approaches [19, 44] based mainly on attention scores, our adaptive strategy overcomes the over-sampling problem effectively. Our framework reduces the time cost from two aspects. First, in contrast to adopting a ﬁxed num- ber of iterations for all cases [12, 36, 38], it runs fewer it- erations for easy cases with few viewpoint or appearance changes and more for challenging cases. Second, it re- duces the cost of each iteration, signiﬁcantly reducing the quadratic time complexity of attention computation. We also show that discarding potential outliers increases not only efﬁciency but also accuracy (see Sec. 5). The efﬁcient version of IMP is called EIMP. Ours contributions are as follows: We propose to perform geometry-aware matching and pose estimation iteratively, allowing the two tasks to boost each other in an iterative manner. We adopt a robust sampling strategy to adaptively dis- card redundant keypoints in the iteration process, sig- niﬁcantly decreasing the time complexity. We apply the pose uncertainty to the sampling strat- egy, which further improves the accuracy matching and pose estimation. Our experiments on relative pose estimation and large- scale localization tasks demonstrate that our method out- performs previous competitors and is more efﬁcient. We organize the rest of the paper as follows. In Sec. 2, we dis- cuss related works. In Sec. 3, we give a detailed descriptionof our method. We test the performance of our model in Sec. 5 and conclude the paper in Sec. 6.
Xie_VideoTrack_Learning_To_Track_Objects_via_Video_Transformer_CVPR_2023	Abstract Existing Siamese tracking methods, which are built on pair-wise matching between two single frames, heavily rely on additional sophisticated mechanism to exploit tempo- ral information among successive video frames, hindering them from efﬁciency and industrial deployments. In this work, we resort to sequence-level target matching that can encode temporal contexts into the spatial features througha neat feedforward video model. Speciﬁcally, we adapt thestandard video transformer architecture to visual trackingby enabling spatiotemporal feature learning directly fromframe-level patch sequences. To better adapt to the track-ing task, we carefully blend the spatiotemporal informationin the video clips through sequential multi-branch triplet blocks, which formulates a video transformer backbone. Our experimental study compares different model variants, such as tokenization strategies, hierarchical structures, and video attention schemes. Then, we propose a disentan- gled dual-template mechanism that decouples static and dy- namic appearance clues over time, and reduces temporalredundancy in video frames. Extensive experiments show that our method, named as VideoTrack, achieves state-of- the-art results while running in real-time.	1. Introduction Visual Object Tracking (VOT) is a fundamental problem in computer vision that aims to track an object of interest ina video given its bounding box in the ﬁrst frame [ 53]. In re- cent years, mainstream approaches formulate visual track- ing as a target matching problem, striking a good balance between performance and simplicity. The philosophy of target matching is to ﬁnd the object by looking for locations in the search area whose features havethe largest similarity with those in the target template. How- *This work was done when Fei Xie was an intern at Microsoft Research Asia. (PEHGGLQJ /D\HU6SDWLRWHPSRUDO )HDWXUH/HDUQLQJ3UHGLFWLRQ 1HWZRUN ݂௭ ݂௧଴ ݂௧ଵ ݂௧௡ ݂௫݂௧௡݂݂ ݂ǥ ݂௫݂݂ 6HTXHQFHOHYHOPDWFKLQJLQ9LGHR7UDFNIUDPHZRUN9LGHRFOLSLQSXW 3UHGLFWLRQ 1HWZRUN (PEHGGLQJ 1HWZRUN/D\HU ݂௫0DWFKLQJ2SHUDWRU 1HWZRUN݂௭ (PEHGGLQJ 1HWZRUN/D\HU7HPSODWH 6HDUFKIUDPH D3DLUZLVHPDWFKLQJLQ6LDPHVHWUDFNLQJIUDPHZRUND 7HPSRUDOFRQWH[WSURSDJDWLRQ2QOLQHWHPSODWHXSGDWLQJ DDD Figure 1. Comparing to the pair-wise matching pipeline in Siamese tracking shown in (a), which requires sophisticated mech-anisms (a1)/(a2) to exploit temporal contexts, our neat video trans- former tracking (VideoTrack) framework, as shown in (b), directly lifts the pair-wise feature matching into spatiotemporal domain. ever, target matching methods generally adopt per-frame object matching manner, where the rich temporal informa- tion in the video is largely overlooked. The representa-tive methods are Siamese trackers [ 4,20,26,27,57]: pair- wise frames are fed into Siamese network to extract fea-tures and a matching network/operator is applied for targetmatching. Although recent pure transformer-based Siamese trackers [ 10,16,55,61] unify feature extraction and match- ing into a single step by leveraging the Vision Transformer(ViT) [ 14,31,49], these Siamese trackers still follow a pair- wise matching philosophy that hinders their exploitation oftemporal context. To explore temporal information, some works have pro- posed sophisticated yet complex temporal modelling meth- ods to improve the robustness of pair-wise Siamese track- ers, where online template updating [ 60,64] and tempo- ral context propagating among frame-wise features [ 47] are two widely-adopted paradigms. Despite their great success, extra hand-crafted hyper-parameters and complex network modules are inevitably introduced to the Siamese pipeline, which have a negative impact on the efﬁciency and are not friendly to embedded devices. A natural question therefore arises: can we exploit the temporal context while still main- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22826 tain the tracking pipeline in a neat, end-to-end fashion? To get rid of the customized updating strategies and re- dundant temporal modelling modules, we directly expandthe pair-wise input frames into video-level (see Fig. 1), to capture rich temporal contexts. Speciﬁcally, we resort tovideo transformers [ 1] to learn spatiotemporal features, and establish the inter-frame temporal dependencies by simplefeedforward modelling. Compared to the popular pair-wiseSiamese matching [ 4,27], our video transformer tracking pipeline (VideoTrack) lifts the 2D matching pipeline into the spatiotemporal domain, allowing the direct processing of video-level inputs. Moreover, we modify the video trans-former architecture to better adapt it to tracking task, based on the following observations and prior knowledge: Feature learning. A good feature representation is vital for the downstream vision tasks. Equipped with dedicated learning scheme in network layers, feature representations can be effectively enhanced from the shallow to deep level. Thus, we attempt to encode the temporal contexts at feature- level by utilizing a video transformer backbone. To ensure the generality and feasibility, we design our video trans- former model with the following principles: 1) Scalabil- ity: Transformer layer is the basic building unit that can be stacked to construct the backbone network in different model scales. 2) Compatibility : To avoid expensive pre- training costs, the modiﬁed network should ideally be com-patible with the model paramaters of the image-based visionbackbone, e.g. ViT [ 14]. It not only can utilize the available pretraining weights, but also prevents the possible perfor-mance degeneration during ﬁne-tuning. Appearance vs. motion clue. Video could be viewed as a temporal evolution of a static appearance. Compared to thevideo recognition task which takes the complete frame as input, the input frames for most trackers are locally cropped from the online predicted target location, which weakens the motion clue in video-clips. Thus, we focus more on uti- lizing the appearance clues. To leverage the strong prior invideo sequences, we explicitly divide them into three cate-gories: initial frame containing strong appearance informa-tion, intermediate frames which contain the dynamic statesof the target and search frame containing the target to be predicted. Thus, we formulate a three-branch architecture for the VideoTrack model. Temporal redundancy. Consecutive video frames are highly redundant. It is vital to reduce the temporal re- dundancy as well as effectively modelling temporal con- texts. Thus, we evaluate three basic temporal modellingapproaches in terms of efﬁciency, i.e. joint space-time, tem- poral window and message token attention. With careful analysis, we propose a disentangled dual-template mecha- nism (see Sec. 3.4for details) to integrate into the video backbone which decouples the redundant video informationinto the static & dynamic templates.As shown in Fig. 2, we propose our VideoTrack frame- work on top of ViT [ 14], formulated by interleaving a series of building units, named as triplet-block. The triplet-blockhas three hierarchical attention layers that mix the informa-tion ﬂow asymmetrically among three branches. Spatiotem-poral guidance from the historical frames is passed to thecurrent search frame, obtaining a compact feature represen-tation for the ﬁnal target prediction. In summary, the main contributions are as follows: • In contrast to existing Siamese tracking methods and their labor-intensive temporal modelling, we for the ﬁrst timelift the 2D pair-wise matching to spatiotemporal domain,encoding temporal context at the feature-level via a neat feedforward video model, i.e. video vision transformer. • We make the ﬁrst attempt to adapt video transformer to visual tracking. A thorough ablation analysis of video transformer tracking is conducted, including tokenisation strategies, model variants and temporal modelling ap- proaches. Comprehensive analysis may inspire follow-ers to solve VOT task from the perspective of video-level modelling. Moreover, our tracker exhibits encouraging results on multiple VOT benchmarks.
Vaze_GeneCIS_A_Benchmark_for_General_Conditional_Image_Similarity_CVPR_2023	Abstract We argue that there are many notions of ‘similarity’ and that models, like humans, should be able to adapt to these dynamically. This contrasts with most representation learn- ing methods, supervised or self-supervised, which learn a ﬁxed embedding function and hence implicitly assume a sin- gle notion of similarity. For instance, models trained on Im- ageNet are biased towards object categories, while a user might prefer the model to focus on colors, textures or spe- ciﬁc elements in the scene. In this paper, we propose the GeneCIS (‘genesis’) benchmark, which measures models’ ability to adapt to a range of similarity conditions. Ex- tending prior work, our benchmark is designed for zero- shot evaluation only, and hence considers an open-set of similarity conditions. We ﬁnd that baselines from powerful CLIP models struggle on GeneCIS and that performance on the benchmark is only weakly correlated with ImageNet ac- curacy, suggesting that simply scaling existing methods is not fruitful. We further propose a simple, scalable solution based on automatically mining information from existing image-caption datasets. We ﬁnd our method offers a sub- stantial boost over the baselines on GeneCIS, and further improves zero-shot performance on related image retrieval benchmarks. In fact, though evaluated zero-shot, our model surpasses state-of-the-art supervised models on MIT-States. We, the architects of the machine, must decide a-priori what constitutes its ‘world’; what things are to be taken as ‘similar’ or ‘equal’ — Karl Popper, 1963	1. Introduction Humans understand many notions of similarity and choose speciﬁc ones depending on the task at hand [ 21,58]. Consider the task of ﬁnding ‘similar’ images illustrated in Figure 1. Which of the rightmost images should be con- sidered ‘most similar’ to the reference? Given different con- ditions , each image could be a valid answer. For instance, we may be interested in a speciﬁc object in the scene, focus- ing on either the ‘car’ or ‘bridge’. One could even indicate *Work done during an internship at Meta AI Research. With the same bridge With a black carWith the same car Figure 1. Given different conditions (shown as blue text), differ- ent images on the right can be considered most ‘similar’ to the reference on the left. We present a general way to train and evalu- ate models which can adapt to different notions of similarity. a ‘negative’ similarity condition, specifying a change in the image to identify the bottom image as most similar. Learning such similarity functions is a central goal in discriminative deep learning [ 11–13,34,63,68,75]. Dis- criminative models, either supervised [ 30,75] or self- supervised [ 9,10], learn embedding functions such that ‘similar’ images are closer in feature space than ‘dissimilar’ images. However, since there are inﬁnitely many notions of image similarity, how do we allow our models to choose? Almost all current approaches assume a single notion of similarity, either by explicitly training on a speciﬁc concept [68,75] or through an implicit assumption in the underlying data distribution [ 9,12]. Meanwhile, prior works tackling the conditional problem have focused on constrained do- mains such as fashion [ 69,73] or birds [ 46], with a restricted set of similarity conditions. This is because developing and evaluating models that can adapt to generic notions of sim- ilarity is extremely challenging. Speciﬁcally, curating data to train and evaluate such models is difﬁcult, as collecting annotations for all concepts of similarity is impossible. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6862 In this work we study the problem of general conditional image similarity, training on an open-set of similarity con- ditions, and evaluating on diverse similarity notions in a ‘zero-shot’ manner. We ﬁrst design a benchmark compris- ing of four evaluation datasets for conditional image simi- larity, setting up conditional retrieval tasks. We deﬁne these tasks under a uniﬁed framework which spans practical use cases, and propose the benchmark as a sparse but broad coverage of the conditional similarity space. We propose these datasets for zero-shot evaluation only , and suggest that models which can perform well without ﬁne-tuning can ﬂexibly adapt to general notions of similarity, as desired. We name this benchmark GeneCIS (‘ genesis ’) for Gene ral Conditional Image Similarity. On GeneCIS, we ﬁnd that baselines built from powerful CLIP backbones struggle and, moreover, that performance on it is only weakly correlated with the backbones’ ImageNet accuracy [ 17]. This is in contrast to popular vision tasks such as segmentation [ 39] and detection [ 45], underlining the benchmark’s utility. We also propose a solution to training general condi- tional similarity models, based on parsing large-scale cap- tion datasets [ 64,66]. Rather than requiring exhaustive sim- ilarity annotations, we ﬁnd that we can automatically mine this information from already abundant image-caption data. We show that training in this way offers substantial gains over the baselines, approaching (and in some cases sur- passing) carefully designed speciﬁc solutions for each of the GeneCIS tasks. In addition, we demonstrate that our method scales with increasing amounts of caption data, sug- gesting promising directions for future work. Finally, on related benchmarks from the ‘Composed Image Retrieval’ (CIR) ﬁeld [ 44,74], we ﬁnd our method provides gains over zero-shot baselines. In fact, our model outperforms state- of-the-art on the MIT-States benchmark [ 28], despite being evaluated zero-shot and never seeing the training data. Contributions. (i)We present a framework for considering conditional image similarity, an important but understudied problem; (ii)We propose the GeneCIS benchmark to test models’ abilities to dynamically adapt to different notions of similarity; (iii)We show that current vision-language models like CLIP struggle on GeneCIS, and that perfor- mance on it is only weakly correlated with ImageNet accu- racy; (iv)We design a scalable solution to the conditional similarity problem based on automatically parsing large- scale image-caption data; (v)We show our models provide substantial gains over zero-shot CLIP baselines; (vi)We validate our models on related CIR benchmarks, surpassing state-of-the-art on MIT-States despite zero-shot evaluation.
Wang_AltFreezing_for_More_General_Video_Face_Forgery_Detection_CVPR_2023	Abstract Existing face forgery detection models try to discriminate fake images by detecting only spatial artifacts ( e.g., gener- ative artifacts, blending) or mainly temporal artifacts ( e.g., flickering, discontinuity). They may experience significant performance degradation when facing out-domain artifacts. In this paper, we propose to capture both spatial and tempo- ral artifacts in one model for face forgery detection. A simple idea is to leverage a spatiotemporal model (3D ConvNet). However, we find that it may easily rely on one type of arti- fact and ignore the other. To address this issue, we present a novel training strategy called AltFreezing for more general face forgery detection. The AltFreezing aims to encourage the model to detect both spatial and temporal artifacts. It divides the weights of a spatiotemporal network into two groups: spatial-related and temporal-related. Then the two groups of weights are alternately frozen during the training process so that the model can learn spatial and temporal features to distinguish real or fake videos. Furthermore, we introduce various video-level data augmentation methods to improve the generalization capability of the forgery detec- tion model. Extensive experiments show that our framework outperforms existing methods in terms of generalization to unseen manipulations and datasets.	1. Introduction With the recent rapid development of face generation and manipulation techniques [30, 31, 45 –49, 56], it has be- come very easy to modify and manipulate the identities or attributes given a face video. This brings many important and impressive applications for movie-making, funny video generation, and so on. However, these techniques can also be abused for malicious purposes, creating serious crisis of trust *Equal contribution. †Corresponding authors. Temporal conv Spatial conv !+ Temporal conv ! Spatial conv+Figure 1. Illustration of AltFreezing training strategy in a build- ing block of the spatiotemporal network. The convolutional ker- nels of the spatiotemporal network are divided into two groups: temporal-based andspatial-based . Two groups of weights are alternately frozen during training. With the help of the alternate freezing ( AltFreezing ) strategy, our model can capture both spatial and temporal artifacts to distinguish between fake and real videos. and security in our society. Therefore, how to detect video face forgeries has become a hot research topic recently. To date, one successful line of research [10, 32, 34, 39, 42, 44, 50] tries to discriminate fake images by detecting “spatial” artifacts in the generated images ( e.g., checkboard, unnaturalness, and characteristic artifacts underlying the generative model). While these methods achieve impressive results in searching spatial-related artifacts, they ignore the temporal coherence of a video and fail to capture “temporal” artifacts like flicking and discontinuity in the video face forgeries. Some recent works [25,43,54] notice this issue and try to address it by leveraging temporal clues. Although they achieve encouraging results in detecting unnatural artifacts at the temporal level, the resulting models are not sufficiently capable of finding spatial-related artifacts. In this paper, we attempt to capture both spatial and tem- poral artifacts for general video face forgery detection. Gen- erally, a well-trained spatiotemporal network (3D ConvNet) has the capability of searching both spatial and temporal artifacts. However, we find that naïve training may cause it to easily rely on spatial artifacts while ignoring temporal artifacts to make a decision, causing a poor generalization This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4129 capability. This is because spatial artifacts are usually more obvious than temporal incoherence, naïvely optimizing a 3D convolutional network makes it easily rely on spatial artifacts. So the question is how to enable the spatiotemporal net- work to capture both spatial and temporal artifacts. To this end, we propose a novel training strategy called AltFreez- ing. As shown in Fig. 1, the key idea is to alternately freeze spatial- and temporal-related weights during training. Specif- ically, a spatiotemporal network [9] is built upon 3D res- blocks, which consist of spatial convolution with kernel size as1×Kh×Kwand temporal convolution with kernel size asKt×1×1. These spatial and temporal convolutional ker- nels are responsible for capturing spatial- and temporal-level features, respectively. Our AltFreezing strategy encourages the two groups of weights to be updated alternately so that both spatial and temporal artifacts can be conquered. Furthermore, we propose a set of video-level fake video argumentation methods for generating fake videos for train- ing. These methods could be divided into two groups. The first is fake clips that only involve temporal artifacts wherein we just randomly drop and repeat frames for real clips. The second is clips with only spatial artifacts that are obtained by blending a region from one real clip to another real clip. These augmentation methods are the first to take the tempo- ral dimension into consideration and generate spatial-only and temporal-only fake videos. With these augmentations, the spatiotemporal model is further encouraged to capture both spatial and temporal artifacts. Equipped with the above-mentioned two techniques, we achieve state-of-the-art performance in various challenging face forgery detection scenarios, including generalization capability to unseen forgeries, and robustness to various perturbations. We also provide a comprehensive analysis of our method to verify the effectiveness of our proposed framework. Our main contributions are three-fold as follows. •We propose to explore both spatial and temporal arti- facts for video face forgery detection. To achieve this, a novel training strategy called AltFreezing is proposed. •We introduce video-level fake data augmentation meth- ods to encourage the model to capture a more general representation of different types of forgeries. •Extensive experiments on five benchmark datasets in- cluding both cross-manipulation and cross-dataset eval- uations demonstrate that the proposed method sets new state-of-the-art performance.
Wang_Robust_Multiview_Point_Cloud_Registration_With_Reliable_Pose_Graph_Initialization_CVPR_2023	Abstract In this paper, we present a new method for the multi- view registration of point cloud. Previous multiview regis- tration methods rely on exhaustive pairwise registration to construct a densely-connected pose graph and apply Itera- tively Reweighted Least Square (IRLS) on the pose graph to compute the scan poses. However, constructing a densely- connected graph is time-consuming and contains lots of outlier edges, which makes the subsequent IRLS struggle to find correct poses. To address the above problems, we first propose to use a neural network to estimate the overlap be- tween scan pairs, which enables us to construct a sparse but reliable pose graph. Then, we design a novel history reweighting function in the IRLS scheme, which has strong robustness to outlier edges on the graph. In comparison with existing multiview registration methods, our method achieves 11% higher registration recall on the 3DMatch dataset and ∼13% lower registration errors on the Scan- Net dataset while reducing ∼70% required pairwise regis- trations. Comprehensive ablation studies are conducted to demonstrate the effectiveness of our designs. The source code is available at https://github.com/WHU- USI3DV/SGHR .	1. Introduction Point cloud registration is a prerequisite for many tasks such as 3D reconstruction [17, 25, 32] and 3D segmenta- tion [27, 35]. Most recent registration methods [1, 7, 22, 28, 39, 45, 48, 56] mainly focus on pairwise registration of two partial point clouds (scans), which can only reconstruct a part of the scene. In order to get a completed scene recon- *Both authors contribute equally to this research. †Corresponding authors: [dongzhenwhu, bshyang]@whu.edu.cn Figure 1. Overview . (1) Given Nunaligned partial scans, our tar- get is to register all these scans into (4) a completed point cloud. Our method has two contributions. (2) We learn a global feature vector to initialize a sparse pose graph which contains much less outliers and reduces the required number of pairwise registrations. (3) We propose a novel IRLS scheme. In our IRLS scheme, we initialize weights from both global features and pairwise registra- tions. Then, we design a history reweighting function to iteratively refine poses, which improves the robustness to outliers. struction, all partial point clouds should be simultaneously aligned, which is called multiview registration . Due to its complexity, multiview point cloud registration receives less attention recently and only few recent studies propose mul- tiview registration methods [18, 21, 30, 53, 54]. Given Nunaligned partial point clouds, multiview reg- istration aims to find a globally-consistent pose for every partial point cloud. A commonly-adopted pipeline of mul- tiview registration consists of two phases [54]. First, a pairwise registration algorithm [28, 45, 48] is applied to ex- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9506 haustively estimate the relative poses of allN 2 scan pairs, which forms a fully-connected pose graph. The edges of the graph stand for the relative poses of scan pairs while nodes represent scans. Since the dense pose graph may include in- accurate or even incorrect relative poses (outliers) between two irrelevant scans, in the second phase, these pairwise poses are jointly optimized by enforcing the cycle consis- tency [30] to reject outlier edges and improve accuracy. For the second phase, most recent methods, including hand- crafted methods [5, 13, 29] or learning-based [21, 30, 54] methods, follow a scheme of Iterative Reweighting Least Square (IRLS). In the IRLS, initial weights are assigned to edges to indicate these edges are reliable or not. Then, based on the weights, a synchronization algorithm is ap- plied to compute a new relative pose on every edge. After that, the weights on edges are updated according to the dif- ference between the old relative poses and the new ones. IRLS iteratively synchronize poses from edge weights and update weights with synchronized poses. In an ideal case, an IRLS scheme will gradually lower the weights of the outlier edges and only consider the in- lier edges for pose synchronization. However, the initial densely-connected graph contains lots of outliers, which of- ten prevents the iterative reweighting mechanism of IRLS from finding correct edges. To improve the robustness to outliers, many researches focus on applying advanced hand- crafted reweighting functions [11, 29] or designing graph network to learn reweighting functions [30, 54]. How- ever, the handcrafted reweighting functions usually require a good initialization to converge to the correct poses while learning-based reweighting methods may not generalize to unseen settings. Designing a robust IRLS algorithm still remains an open problem. In this paper, we show that multiview registration can be improved from two aspects, as shown in Fig. 1. First, we learn a good initialization of the input pose graph which avoids exhaustive pairwise registrations and reduces the outlier ratio. Second, we propose a novel history reweight- ing function which enables a stable convergence to correct poses in the IRLS scheme. In the pose graph construction, we learn a global feature on each point cloud and the correlation of two global feature indicates the overlap ratio between two point clouds. Such global features enable us to generate a sparse pose graph with fewer but more reliable edges instead of a densely- connected graph. After that, we only need to apply the pairwise registration algorithm and IRLS on these sparse edges, which greatly reduce the computation complexity of pairwise registration from O(N2)toO(N). Mean- while, these reliable edges contain much less outliers than the fully-connected graph, which provides the possibility to find more accurate and consistent global poses in IRLS. Though the initial graph contains much less outliers, ex- Figure 2. An example on the 3DMatch dataset. (a) The input scans under the ground truth poses. (b) The constructed sparse pose graph with two incorrect relative poses (#0-#2 and #0-#4), where #0 and #4 looks very similar to each other so that the pose graph incorrectly include this scan pair. (c) and (d) show the nor- malized weights on the graph edges on different iterations of the vanilla IRLS and our method respectively. Our method is able to find the outlier edges and gradually reduce their weights while vanilla IRLS is biased towards the outlier edge (#0-#4) after few iterations. (e) and (f) are the multiview registration results of the vanilla IRLS and our method respectively. isting IRLS algorithms are still sensitive to these outliers and can be totally biased towards these outliers in the first few iterations. An example is shown in Fig. 2: the initial graph only contains two outlier edges. However, the outlier scan pair “#0-#4” looks very similar and thus is initialized with a large weight. Such an incorrect large weight inter- feres the subsequent pose synchronization and brings sys- tematic errors to the synchronized poses. The vanilla IRLS trusts all synchronized poses and is easily dominated by these erroneous poses, which leads to incorrect convergence as shown in Fig. 1(c). To address this problem, we propose a simple yet effective reweighting function called the his- tory reweighting function . In history reweighting function, edge weights at a specific iteration not only depends on the synchronized poses at the current iterations but also consid- ers the historical synchronized poses in previous iterations, which acts like a regularizer to prevent the IRLS from be- ing dominated by outliers at the early unstable iterations as shown in the Fig. 2 (d). Then, the edge weights in our graph gradually stabilize in the subsequent iterative refinements, leading to the convergence to correct poses. 9507 We evaluate our method on three widely-used bench- marks: the 3DMatch/3DLoMatch dataset [28,58], the Scan- Net dataset [16], and the ETH dataset [43]. With the help of the proposed sparse graph construction and IRLS with history reweighting, our method surpasses the current mul- tiview registration baselines by 11.0%and6.2%in reg- istration recall on 3DMatch and 3DLoMatch, reduces the mean rotation and translation errors on ScanNet by 12.8% and13.8%. Meanwhile, our method shows strong gener- alization ability. Only trained on the indoor dataset, our method achieves a 99.8%registration recall on the outdoor ETH dataset. Moreover, all the above state-of-the-art per- formances of our method only require 20%∼40% pairwise registrations of existing multiview point cloud registration methods, which demonstrates our computation efficiency.
Xu_Gaussian_Label_Distribution_Learning_for_Spherical_Image_Object_Detection_CVPR_2023	Abstract Spherical image object detection emerges in many appli- cations from virtual reality to robotics and automatic driv- ing, while many existing detectors use ln-norms loss for re- gression of spherical bounding boxes. There are two intrin- sic flaws for ln-norms loss, i.e., independent optimization of parameters and inconsistency between metric (dominated by IoU) and loss. These problems are common in planar image detection but more significant in spherical image de- tection. Solution for these problems has been extensively discussed in planar image detection by using IoU loss and related variants. However, these solutions cannot be mi- grated to spherical image object detection due to the undif- ferentiable of the Spherical IoU (SphIoU). In this paper, we design a simple but effective regression loss based on Gaus- sian Label Distribution Learning (GLDL) for spherical im- age object detection. Besides, we observe that the scale of the object in a spherical image varies greatly. The huge differences among objects from different categories make the sample selection strategy based on SphIoU challeng- ing. Therefore, we propose GLDL-ATSS as a better training sample selection strategy for objects of the spherical image, which can alleviate the drawback of IoU threshold-based strategy of scale-sample imbalance. Extensive results on various two datasets with different baseline detectors show the effectiveness of our approach.	1. Introduction In the past few years, with the numerous development of panoramic cameras with omnidirectional vision, the ap- plications of spherical images and videos are also becom- ing more extensive, such as virtual & augmented real- ity [9,17,24], robotics [7,8,18], automatic driving [1,28,31], etc. As these spherical data increase, the demand for spher- *This work was done when Hang Xu and Qiang Zhao were at ICT. †Corresponding author. (a) (b) (c) (a) Planar imageTop: \|\|.\|\| 1 = 8.42 IoU = 0.36 Bottom: \|\|.\|\| 1 = 8.42 IoU = 0.36Top: \|\|.\|\| 1 = 6.78 IoU = 0.38 Bottom: \|\|.\|\| 1 = 6.78 IoU = 0.24 (b) Spherical image Figure 1. Comparison between planar image and spherical image. (a) Moving centers of two bounding boxes in the planar image along the y-axis does not change the distance between the two centers, so IoU and L1 are unchanged. (b) Moving centers of two bounding boxes to the equator along the longitude changes the distance between two centers, which causes IoU decrease sharply whereas the L1 value is unchanged. ical vision analysis tasks increases, especially the object detection task of spherical image. However, compared with the large literature on planar image object detection [2, 12, 16, 34, 39], research in spherical image object detec- tion is relatively in its earlier stage, with many open prob- lems to solve. In spherical image object detection, a bounding box is represented by a Bounding Field of View (BFoV) [25]. Many existing detection benchmarks [4, 5, 26, 27, 29] use ln-norms loss for the regression of BFoVs. However, the ln-norms loss has some intrinsic flaws. First, parameters of the bounding box are optimized independently in the ln- norms loss, leading to the detection accuracy sensitive to the fitting of any of the parameters. Second, Intersection This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1033 Detection HeadRBFoVLabel Information Backbone with Feature PyramidGaussian Distribution Sample Sele ction Regression LossZ XYP O Z XOP YZ XOYP1P2+ Statistical Distance:Figure 2. Overview of our main contributions. Gaussian distributions of spherical bounding boxes are constructed, and the sample selection strategy (GLDL-ATSS) and regression loss (GLDL loss) are designed in an alignment manner on the basis of K-L divergence. Note that the GLDL-ATSS and GLDL loss are not involved in the inference phase. Therefore, the inference time remains unchanged. over Union (IoU) has been the standard metric for object detection, so ln-norms as regression loss cause the negative impact of the inconsistency between metric and loss. These problems are common in planar image detection but more significant in spherical image detection. Fig. 1(b) shows the inconsistency between IoU and L1 Loss in the spherical im- age. Specifically, moving centers of bounding boxes to the equator along the longitude changes the distance between two centers, which causes IoU decrease sharply while the L1 value is unchanged. In contrast, as shown in Fig. 1(a), moving centers of bounding boxes in the planar image along the y-axis does not change the distance between the two centers, so IoU and L1 are unchanged. Solutions for these problems of ln-norms loss have become recently popular in planar image detection by using IoU-induced loss, such as IoU loss [32], GIoU loss [23] and DIoU loss [37]. How- ever, when calculating the intersection area of two spher- ical boxes, the number of intersection points needs to be obtained. When two spherical boxes are completely coin- cident or one edge is coincident, the number of intersection points will not be fixed and duplicate points will appear. The current SphIoU uses the DFS algorithm to remove these duplicate intersection points. To the best of our knowledge, the DFS algorithm is undifferentiable. Therefore, Spherical IoU (SphIoU) [5,27] is undifferentiable and these solutions based on IoU loss cannot be migrated to spherical image ob- ject detection. The more recent work [30] finds the key to maintaining the consistency between metric and regression loss lies in the trend-level consistency between regression loss and IoU loss rather than value-level consistency, which greatly decreases the difficulty of designing alternatives. In this paper, we design a simple but effective regres- sion loss based on Gaussian Label Distribution Learning (GLDL) for spherical image object detection. Specifically, in the training phase, we first convert tangent planes of the predicted spherical bounding box and ground truth box intothe Gaussian distribution. Then, we devise a dynamic sam- ple selection strategy (GLDL-ATSS) to select positive sam- ples, which can alleviate the drawback of IoU threshold- based strategy of scale-sample imbalance. Finally, we de- sign a regression loss function based on GLDL for spherical object detection task. We observe that GLDL loss achieves a trend-level alignment with SphIoU loss. In the inference phase, we directly obtain the output for the spherical bound- ing box from the trained model of the parameter weights, so the inference time of the network remains unchanged. The entire framework of the method in this paper is shown in Fig. 2. The highlights of this paper are as follows: • We explore a new regression loss function based on Gaussian Label Distribution Learning (GLDL) for spherical object detection task. It achieves a trend- level alignment with SphIoU loss and thus naturally improves the model. • We align the measurement between sample selection and loss regression based on the GLDL, and then construct new dynamic sample selection strategies (GLDL-ATSS) accordingly. GLDL-ATSS can allevi- ate the drawback of IoU threshold-based strategy (i.e., scale-sample imbalance). • Extensive experimental results on two datasets and popular spherical image detectors show the effective- ness of our approach.
Xue_Stare_at_What_You_See_Masked_Image_Modeling_Without_Reconstruction_CVPR_2023	Abstract Masked Autoencoders (MAE) have been prevailing paradigms for large-scale vision representation pre- training. By reconstructing masked image patches from a small portion of visible image regions, MAE forces the model to infer semantic correlation within an image. Re- cently, some approaches apply semantic-rich teacher mod- els to extract image features as the reconstruction target, leading to better performance. However, unlike the low- level features such as pixel values, we argue the features extracted by powerful teacher models already encode rich semantic correlation across regions in an intact image. This raises one question: is reconstruction necessary in Masked Image Modeling (MIM) with a teacher model? In this paper, we propose an efficient MIM paradigm named MaskAlign. MaskAlign simply learns the consistency of visible patch features extracted by the student model and intact image features extracted by the teacher model. To further ad- vance the performance and tackle the problem of input in- consistency between the student and teacher model, we pro- pose a Dynamic Alignment (DA) module to apply learn- able alignment. Our experimental results demonstrate that masked modeling does not lose effectiveness even with- out reconstruction on masked regions. Combined with Dynamic Alignment, MaskAlign can achieve state-of-the- art performance with much higher efficiency. Code and models will be available at https://github.com/ OpenPerceptionX/maskalign .	1. Introduction In recent years, Vision Transformers are showing tremendous potential in computer vision area [8, 40, 41]. Following the big success of masked modeling in the nat- ural language processing [24], Masked Image Modeling *This work was performed when Hongwei Xue was visiting Shanghai AI Laboratory as a research intern. †Corresponding authors. Encoder EncoderEncoderDecoder Visible Masked Target(a) (b) (c)Add Position Embedding HeadFigure 1. Comparison with existing paradigms of masked im- age modeling. (a) Inpainting-style : BEiT [3], MaskFeat [43], MVP [44], BEiT V2 [33], etc. They take the whole image with some mask token replacement as the input of Encoder. Then a Linear head is applied to predict masked feature. (b) Decoder- style : MAE [13], CAE [5], MCMAE [11], etc. They drop most tokens and take the rest as the input of Encoder. Then a multi-layer Transformer is applied to decode masked features from visible to- kens. (c) Ours. : Our paradigm take some visible tokens as the input of Encoder and align visible tokens with target features only. (MIM) has demonstrated a great ability of self-supervised learning [3, 13], while alleviating the data-hungry issue of Transformer architectures. The visual representation learned through MIM shows promising performance on various downstream vision tasks, outperforming the con- trastive learning paradigms [4, 6]. Existing Masked Image Modeling (MIM) methods aim to hallucinate the intact image from a small portion of vis- ible image regions. As depicted in Fig. 1, existing MIM methods are mainly divided into two types: (a) inpainting- style [3, 43, 46] and (b) decoder-style [5, 11, 13]. These two types both require the model to reconstruct masked regions. The inpainting-style models replace image regions with learnable vectors then fill them by the interaction within the encoder. The decoder-style models drop image regions then decode features from masked regions’ positions based on This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 22732 the visible information. Some very recent works introduce semantic-rich teacher models like CLIP [35] into the two paradigms by using features extracted by teacher models as the reconstruction target [17, 33, 34, 44]. In light of the se- mantic knowledge learned by teacher models, these works further improve the representation after masked image mod- eling, leading to better performance. Reconstruction on masked regions implicitly forces the model’s encoder to understand the semantic correlations within an image. However, the reconstruction manner brings much computation on masked tokens within or out- side the encoder in inpainting-style or decoder-style, re- spectively. This redundant computation decreases the train- ing efficiency of the encoder thus increasing the pre-training cost. Unlike low-level and isolated features such as normal- ized pixel values of patches, Histogram of Oriented Gra- dients (HOG), etc., the feature map extracted by powerful teacher models already contains rich semantic correlations, learned during the teacher model training stage. This dif- ference raises one question: is reconstruction the only way in Masked Image Modeling (MIM) with teacher models? To answer this question, we propose a much more efficient MIM paradigm named MaskAlign without any reconstruc- tion on masked tokens. On contrary of applying reconstruction on masked to- kens, MaskAlign simply aligns the visible features ex- tracted by the student model and intact image features ex- tracted by the teacher model. As a consequence, MaskAlign forces the student model to learn not only good representa- tion of the teacher model by feature alignment, but also the ability to hallucinate by masked modeling: feature consis- tency between the intact image and mask view requires the student model to infer semantics from much less informa- tion than teacher model. We adopt multi-level features of the teacher model as supervision to borrow richer seman- tics. However, the input of the student model contains much less information than the teacher model’s, leading to mis- alignment of each layer’s features. To tackle this problem, we enhance the student’s features with a Dynamic Align- ment (DA) module. DA dynamically aggregates different levels of student features and aligns with multi-level fea- tures of the teacher model. This approach can also easily transfer to asymmetric student-teacher structures. From our experimental results, MaskAlign with a wide range of mask ratio outperforms the mask ratio of 0%, where it degenerates into Feature Distillation [45]. This verifies that masked modeling is still necessary for our paradigm. Meanwhile, our experiments validate the ef- fectiveness of Dynamic Alignment by comparing different alignment strategies and numbers of feature levels. The combination of masked modeling and Dynamic Alignment makes our model achieve state-of-the-art results with much higher efficiency. For example, our model outperformsBEiT v2 [33] by 0.4% on ImageNet Finetuning Accuracy (from 85.0% to 85.4%) with 1/3 pre-training time only. To sum up, our work has three-fold contributions: 1. We categorize and rethink existing Masked Image Modeling (MIM) paradigms and propose a more ef- ficient MIM approach called MaskAlign. Even with- outany reconstruction on masked tokens, MaskAlign achieves new state-of-the-art performance with much higher efficiency. 2. We propose a Dynamic Alignment (DA) module to tackle the problem of input inconsistency between the student and teacher model, with negligible additional parameters and computation. 3. We conduct extensive experiments to verify the effec- tiveness of MaskAlign and Dynamic Alignment. Be- sides, our model shows a good ability of generalization on downstream tasks and larger size models.
Wang_All_in_One_Exploring_Unified_Video-Language_Pre-Training_CVPR_2023	Abstract Mainstream Video-Language Pre-training (VLP) mod- els [10, 26, 64] consist of three parts, a video encoder, a text encoder, and a video-text fusion Transformer. They pursue better performance via utilizing heavier unimodal encoders or multimodal fusion Transformers, resulting in increased parameters with lower efficiency in downstream tasks. In this work, we for the first time introduce an end- to-end VLP model, namely all-in-one Transformer, that em- beds raw video and textual signals into joint representations using a unified backbone architecture. We argue that the unique temporal information of video data turns out to be a key barrier hindering the design of a modality-agnostic Transformer. To overcome the challenge, we introduce a novel and effective token rolling operation to encode tem- poral representations from video clips in a non-parametric manner. The careful design enables the representation learning of both video-text multimodal inputs and unimodal inputs using a unified model. Our pre-trained all-in-one Transformer is transferred to various downstream video- text tasks after fine-tuning, including text-video retrieval, video-question answering, multiple choice and video cap- tioning. State-of-the-art performances with the minimal model FLOPs on ten datasets demonstrate the superior- ity of our method compared to the competitive counter- parts. The code and pretrained models are available at https://github.com/showlab/all-in-one .	1. Introduction Science advances rather steadily for most of the time, but sometimes has a disruptive episode, where “an older paradigm is replaced in whole or in part by an incompat- ible new one.” [24] In this regard, Video-Language Pre- training (VLP) models have recently experienced steady progress, where joint representations are generally pro- duced with a multimodal fusion network after extracting *Corresponding Author. Figure 1. Compare to mainstream video-language pre-training methods. (a). Conventional methods [3, 10, 26, 64] use deep fea- tures from separate encoders before fusion. The fusion layer can be light [3] or heavy [10,26,64]. (b). Ours All-in-one Transformer learns video and text joint representations end-to-end from their raw inputs. We also support fast retrieval by feeding unimodal in- puts during inference. (c). Comparison of FLOPs and retrieval performance on MSRVTT [53]. Our All-in-one brings excellent results with modest computational cost. the visual and language features through unimodal encoders [10,26,50,64]. We are here to break it and replace them with “an incompatible new one” that has NO unimodal encoders. The pre-train and then fine-tune scheme, has become a standard paradigm to learn transferable video-language representations for a wide range of downstream video-text tasks [6, 15, 52, 52, 53, 61]. Mainstream methods attempt to boost the pre-training in two ways: i. adopting more expen- sive video/text encoders to obtain more powerful unimodal features [3, 10] ii. designing heavier fusion networks to en- hance the association between modalities [61, 64]. Instead of following these trends, we fundamentally re- think design decisions and develop the simplest and most lightweight architecture that learns video-language repre- sentations from their raw inputs in an end-to-end manner. Our model does not need any unimodal encoders ( e.g., ob- ject detector in [64] or ResNet visual encoder in [26]) or complex fusion layers, but embeds visual and text signals in a unified manner, termed as All-in-one Transformer in our paper. Our design is inspired by recent studies [1, 21, 37] This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6598 that perform multimodal pre-training under the presump- tion that Transformer can process visual data in the same way as it processes text. However, our work is not the straightforward application of them. It is not trivial how to embed videos for our unified Transformer due to the unique challenge of modeling temporal information without adding much computational cost. Existing works model temporal information by de- signing temporal attention layers [3] or using temporal- aware visual encoders ( e.g., 3D convnets in [64] or Video Swin [36] in [10]). We cannot simply use them in our unified All-in-one Transformer because they are modality- dependent and computationally too expensive. To address this issue, we design a novel, effective, and efficient method to model temporal information. Our model only needs three frames per video clip, which is much lower than other mod- els (e.g., 16 [26] or 32 [1]) but can achieve the comparable performance to them. Nevertheless, we are still not satis- fied with the computational cost in the self-attention layer. To further reduce the computational cost, we propose the temporal token rolling operation , which is a cyclic atten- tion between small proportions of the visual tokens in each frame (Fig. 3-right). This is much more efficient than a naive self-attention approach on flattened tokens (Fig. 3- bottom). Furthermore, our modality-agnostic design en- ables us to use our pre-trained model as a powerful uni- modal feature extractor by feeding only video or text in- puts. This can significantly reduce the computational cost for retrieval task because we can simply compute the co- sine similarity of texts and videos soon after the pretraining, eliminating the need for training additional fusion module of projecting the disjoint text and visual features into a com- mon space (Fig. 1-b). Taken together, our All-in-one archi- tecture achieves much less FLOPs and better text-to-video performance than previous work (Fig. 1-c), despite the fact that we use the same pre-training objectives [10, 26]. Contributions. (1) We introduce the simplest, most lightweight, and most efficient video-language model, namely All-in-one Transformer, which is the first to capture video-language representations from the raw visual and tex- tual signals end-to-end in a unified backbone architecture. (2) We elucidate and tackle the difficulties of applying a unified and shared backbone for multimodal video and text data, that is, how to properly process the unique temporal information of videos. A novel temporal token rolling op- eration is proposed to capture the temporal representations of sparsely sampled frames without any extra parameters or increasing time complexity. (3) We propose a success prac- tical to overcome the slow retrieval of one-stream model and explore how to cotrain the image and video data together in better ways. (4) Comprehensive experiments on five down- stream video-text tasks of eleven datasets fully demonstrate the superiority of our pre-trained All-in-one Transformer onboth effectiveness and efficiency compared to recent main- stream methods [3, 10, 26].
Wang_Neural_Koopman_Pooling_Control-Inspired_Temporal_Dynamics_Encoding_for_Skeleton-Based_Action_CVPR_2023	Abstract Skeleton-based human action recognition is becoming increasingly important in a variety of fields. Most exist- ing works train a CNN or GCN based backbone to extract spatial-temporal features, and use temporal average/max pooling to aggregate the information. However, these pool- ing methods fail to capture high-order dynamics informa- tion. To address the problem, we propose a plug-and- play module called Koopman pooling, which is a param- eterized high-order pooling technique based on Koopman theory. The Koopman operator linearizes a non-linear dy- namics system, thus providing a way to represent the com- plex system through the dynamics matrix, which can be used for classification. We also propose an eigenvalue nor- malization method to encourage the learned dynamics to be non-decaying and stable. Besides, we also show that our Koopman pooling framework can be easily extended to one-shot action recognition when combined with Dynamic Mode Decomposition. The proposed method is evaluated on three benchmark datasets, namely NTU RGB+D 60, 120 and NW-UCLA. Our experiments clearly demonstrate that Koopman pooling significantly improves the performance under both full-dataset and one-shot settings.	1. Introduction Skeleton-based human action recognition is a crucial task in many applications, ranging from video surveillance to autonomous driving and human-robot interaction. With the prevalence of deep learning, the rising of LSTM, CNN and GCN has significantly improved the performance of ac- tion recognition. Most existing methods [9, 13,28,46,70,73] use a CNN or GCN based backbone to extract complex spatial-temporal features, and use temporal average/max pooling to aggregate the information. However, vanilla tem- poral average/max pooling contains only first-order infor- mation and abandons higher-order statistical information. *Corresponding authorFeature Extraction Aggregation & Classification … 𝑥𝑥1𝑥𝑥2 𝑥𝑥𝑇𝑇 ҧ𝑥𝑥Koopman Space𝑥𝑥1 𝑥𝑥2 𝑥𝑥𝑇𝑇 Class -wise Dynamics Matrices𝐾𝐾2𝐾𝐾1 𝐾𝐾𝑁𝑁𝐾𝐾 𝐾𝐾 Matching… Temporal Average Pooling Average/Sum Koopman PoolingFeature Space Classifier… Figure 1. Most existing works use temporal average pooling to aggregate the information along the temporal dimension (left), and only first-order information is considered. Our proposed Koopman pooling (right) instead focuses on the true latent dynamics of the sequence in linear Koopman space, and learns a set of class-wise Koopman dynamics matrices to represent the dynamics of each class. The classification is achieved by dynamics matching. To this end, recent works focus on second-order pool- ing to capture second-order information of the feature se- quences. Specifically, bilinear and covariance pooling is widely used to extract second-order statistical information from the feature sequence. Early works [14] use sim- ple bilinear pooling to model the interaction of features and aggregate temporal information. In recent years, re- searchers proposed to use covariance pooling [22] to cap- ture second-order statistical information, as covariance ma- trix can model the interaction between features while main- taining good geometrical structure [31]. However, the skeleton sequence as well as the extracted feature sequence has complex underlying dynamics in nature. Existing meth- ods such as covariance pooling only exploit the feature in- teraction between frames or channels, but they fail to dis- cover the true dynamics of the sequence. Instead, we aim This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10597 to directly focus on the temporal dynamics of the sequence and conduct sequence recognition based on class-specific dynamics. A critical motivating idea of this work is the applica- tion of Koopman theory [26]. The original Koopman the- ory aims to re-formulate a non-linear dynamical system to be a linear one. To this end, a Koopman operator is required to lift the original features into some possibly infinite-dimensional Hilbert space, wherein the evolution of the dynamics becomes linear. Such a treatment is fa- vored in numerous applications, including time-series anal- ysis, since an arsenal of spectral-analysis tools can facilitate the in-depth investigation of the model (such as the tempo- ral stability property). In practice, identifying the optimal Koopman operator of a specific dynamic system remains a challenge. Besides conventional dynamic mode decompo- sition (DMD) [6], recent years also witnessed the utilization of black-box neural network for learning the Koopman op- erator in differentiable fashion [1, 23,38,42,61]. To our best knowledge, the proposed Koopman pooling in this paper is the first work to leverage the power of Koop- man theory to formulate a new high-order pooling method. Unlike existing methods like covariance pooling which use covariance matrix to model the temporal correlation of fea- tures implicitly, we instead view the temporal evolution of feature vectors as a dynamical system, and use the dynam- ics itself to model the temporal correlation explicitly. As shown in figure 1, the original trajectories are mapped to a new embedding space where the temporal evolution is lin- ear. The transition matrix of this linear system can therefore be viewed as the signature of this sequence, which contains rich high-order information. For the classification task, our model learns the class-wise Koopman matrices which represent class-specific dynamics, and conducts dynamics matching to obtain the classification score. Based on our observation of the learned dynamics, this paper highlight the critical importance of the stability of learned dynamics when tackling recognition tasks, and propose an eigenvalue normalization technique to push the learned linear dynam- ics to be stable and non-decaying. We also combine Koop- man pooling and dynamic mode decomposition(DMD) to formulate a new framework for one-shot action recognition, which uses dynamics to match the sequence instead of the common practice of computing distance in the embedding space or conducting metric learning. To verify the effectiveness of the proposed method, we conduct extensive experiments on 3 skeleton-based action recognition datasets, namely NTU RGB+D, NTU RGB+D 120, and NW-UCLA. The results demonstrate that Koop- man pooling significantly improves the performance under both full-dataset and one-shot settings. To be summarized, the main contributions of this paper are as follows:•We proposed Koopman pooling, which is the first work in the literature to design a plug-and-play high-order pooling method based on Koopman theory that allows eigenvalue manipulation and one-shot recognition. •We emphasize the critical importance of learning a sta- ble and non-decaying system for recognition tasks and accordingly design an eigenvalue normalization tech- nique based on control theory. •Our comprehensive experiments based on various backbones [9, 28,68] on the commonly used bench- mark datasets NTU RGB+D 60/120 and NW-UCLA shows Koopman pooling significantly improves the performance under both full-dataset and one-shot set- ting.
Wang_Neural_Fields_Meet_Explicit_Geometric_Representations_for_Inverse_Rendering_of_CVPR_2023	Abstract Reconstruction and intrinsic decomposition of scenes from captured imagery would enable many applications such as relighting and virtual object insertion. Recent NeRF based methods achieve impressive ﬁdelity of 3D reconstruc- tion, but bake the lighting and shadows into the radiance ﬁeld, while mesh-based methods that facilitate intrinsic de- composition through differentiable rendering have not yet scaled to the complexity and scale of outdoor scenes. We present a novel inverse rendering framework for large urban scenes capable of jointly reconstructing the scene geometry, spatially-varying materials, and HDR lighting from a set of posed RGB images with optional depth. Speciﬁcally, we use a neural ﬁeld to account for the primary rays, and use an explicit mesh (reconstructed from the underlying neural ﬁeld) for modeling secondary rays that produce higher-order lighting effects such as cast shadows. By faithfully disentan- gling complex geometry and materials from lighting effects, our method enables photorealistic relighting with specular and shadow effects on several outdoor datasets. Moreover, it supports physics-based scene manipulations such as virtual object insertion with ray-traced shadow casting.	1. Introduction Reconstructing high ﬁdelity 3D scenes from captured im- agery is an important utility of scaleable 3D content creation. However, for the reconstructed environments to serve as “dig- ital twins” for downstream applications such as augmented reality and gaming, we require that these environments are compatible with modern graphics pipeline and can be ren- dered with user-speciﬁed lighting. This means that we not only need to reconstruct 3D geometry and texture but also recover the intrinsic properties of the scene such as material properties and lighting information. This is an ill-posed, challenging problem oftentimes referred to as inverse render- ing [1]. Neural radiance ﬁelds (NeRFs) [34] have recently emerged as a powerful neural reconstruction approach that enables photo-realistic novel-view synthesis. NeRFs can be reconstructed from a set of posed camera images in a matter of minutes [14, 35, 43] and have been shown to scale to room-level scenes and beyond [45, 48, 55], making them an attractive representation for augmented/virtual reality and generation of digital twins. However, in NeRF, the intrinsic properties of the scene are not separated from the effect of in- cident light. As a result, novel views can only be synthesised This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 8370 under ﬁxed lighting conditions present in the input images, i.e. a NeRF cannot be relighted [42]. While NeRF can be extended into a full inverse rendering formulation [3], this requires computing the volume render- ing integral when tracing multiple ray bounces. This quickly becomes intractable due to the underlying volumetric rep- resentation. Speciﬁcally, in order to estimate the secondary rays, the volumetric density ﬁeld of NeRF would have to be queried along the path from each surface point to all the light sources, scaling with O(nm)per point, where nde- notes the number of samples along each ray and mis the number of light sources or Monte Carlo (MC) samples in the case of global illumination. To restrict the incurred computa- tional cost, prior works have mostly focused on the single object setting and often assume a single (known) illumina- tion source [42]. Additionally, they forgo the volumetric rendering of secondary rays and instead approximate the direct/indirect lighting through a visibility MLP [42, 64]. In contrast to NeRF, the explicit mesh-based representa- tion allows for very efﬁcient rendering. With a known mesh topology, the estimation of both primary and secondary rays is carried out using ray-mesh intersection ( O(m)) queries that can be efﬁciently computed using highly-optimized li- braries such as OptiX [39]. However, inverse rendering methods based on explicit mesh representations either as- sume a ﬁxed mesh topology [13], or recover the surface mesh via an SDF deﬁned on a volumetric grid [36] and are thus bounded by the grid resolution. Insofar, these methods have been shown to produce high-quality results only for the smaller, object-centric scenes. In this work, we combine the advantages of the neural ﬁeld (NeRF) and explicit (mesh) representations and pro- pose FEGR1, a new hybrid-rendering pipeline for inverse rendering of large urban scenes. Speciﬁcally, we represent the intrinsic properties of the scene using a neural ﬁeld and estimate the primary rays (G-buffer) with volumetric render- ing. To model the secondary rays that produce higher-order lighting effects such as specular highlights and cast shad- ows, we convert the neural ﬁeld to an explicit representa- tion and preform physics-based rendering. The underlying neural ﬁeld enables us to represent high-resolution details, while ray tracing secondary rays using the explicit mesh reduces the computational complexity. The proposed hybrid- rendering is fully differentiable an can be embedded into an optimization scheme that allows us to estimate 3D spatially- varying material properties, geometry, and HDR lighting of the scene from a set of posed camera images2. By modeling the HDR properties of the scene, our representation is also well suited for AR applications such as virtual object inser- 1Abbreviation FEGR is derived from neural Fields meet Explicit Geometric Representations and is pronounced as "ﬁgure" . 2We can also integrate depth information, if available, to further con- strain the solution space.tion that require spatially-varying lighting to cast shadows in a physically correct way. We summarize our contributions as follows: We propose a novel neural ﬁeld representation that decom- poses scene into geometry, spatially varying materials, and HDR lighting. To achieve efﬁcient ray-tracing within a neural scene rep- resentation, we introduce a hybrid renderer that renders primary rays through volumetric rendering, and models the secondary rays using physics-based rendering. This en- ables high-quality inverse rendering of large urban scenes. We model the HDR lighting and material properties of the scene, making our representation well suited for down- stream applications such as relighting and virtual object insertion with cast shadows. FEGR signiﬁcantly outperforms state-of-the-art in terms of novel-view synthesis under varying lighting conditions on the NeRF-OSR dataset [40]. We also show qualitative results on a single-illumination capture of an urban environment, collected by an autonomous vehicle. Moreover, we show the intrinsic rendering results, and showcase virtual object insertion as an application. Finally, we conduct a user study, in which the results of our method are signiﬁcantly preferred to those of the baselines.
Xiao_Endpoints_Weight_Fusion_for_Class_Incremental_Semantic_Segmentation_CVPR_2023	Abstract Class incremental semantic segmentation (CISS) fo- cuses on alleviating catastrophic forgetting to improve dis- crimination. Previous work mainly exploits regularization (e.g., knowledge distillation) to maintain previous knowl- edge in the current model. However, distillation alone of- ten yields limited gain to the model since only the repre- sentations of old and new models are restricted to be con- sistent. In this paper, we propose a simple yet effective method to obtain a model with a strong memory of old knowledge, named Endpoints Weight Fusion (EWF). In our method, the model containing old knowledge is fused with the model retaining new knowledge in a dynamic fusion manner, strengthening the memory of old classes in ever- changing distributions. In addition, we analyze the rela- tionship between our fusion strategy and a popular mov- ing average technique EMA, which reveals why our method is more suitable for class-incremental learning. To facili- tate parameter fusion with closer distance in the parameter space, we use distillation to enhance the optimization pro- cess. Furthermore, we conduct experiments on two widely used datasets, achieving state-of-the-art performance.	1. Introduction As a fundamental task, semantic segmentation plays a key role in visual applications [10, 25]. Previous fully- supervised works aim to segment fixed classes defined in the training set. However, the trained segmentation model is expected to recognize more classes in realistic applica- tions. One straightforward solution is to re-train the model on the entire dataset by mixing old and new data. Never- theless, this strategy will bring huge labeling and training costs. From the transfer learning perspective [22, 30], an- other plain solution is to adjust the previously learned model on the newly added data. But the model will overfit to new *The first two authors contribute equally. †Corresponding author (xialei@nankai.edu.cn) CNN (old)CNN2 CNN1 CNNn... ... output2 EnsembleCNN (frozen) CNN (compressed)CompressionCNN (new) frozen branchtrained branch merged branchRe-parameterizationoutput1 outputn FC FC FC non-linear non-linearoutput 1-CNN (new) CNN (balanced)EWF outputFigure 1. Illustration of different fusion strategies for incremen- tal learning. Ensemble methods utilize multiple models to accu- mulate more knowledge. Compression methods reduce the model size and distill the knowledge into a small network. While Re- parameterization methods use equivalent operations for model fu- sion. Our Endpoints Weight Fusion ( EWF ) proposes model addi- tion with a dynamic factor ( αt) with no further training. classes quickly, while forgetting previous old classes. This phenomenon is also known as catastrophic forgetting [35]. To alleviate the problem of catastrophic forgetting with- out extra labeling or training cost, class incremental se- mantic segmentation (CISS) [3, 16, 50] aims at optimiz- ing the trade-off between maintaining discrimination for old classes and learning knowledge of new classes. Most works [3, 16, 17, 37] designed regularization methods to maintain a balance between memorizing old knowledge and learning new one. We observe that existing works can still suffer from catastrophic forgetting, resulting in a significant This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7204 performance drop in old classes. In the scenario of CISS, not only the previous data is not accessible due to privacy issues or data storage limitations, but regions of old classes in the newly added dataset are labeled as background, which further exacerbates the model over-fitting. Besides, training a new model from the old one and fus- ing them to obtain the final model is a common strategy in continual learning. As shown in Fig. 1, we roughly di- vide them into four categories with two stages of model ex- pansion and fusion. Some methods [27, 33, 42, 47] propose to expand the model in incremental steps and ensemble the old and new outputs, which have large memory and infer- ence costs. While some works apply compression [46, 47] to compress the old and new model to a unified model with fewer parameters. Nevertheless, these require further train- ing on only new data, which can lead to a bias toward new data. Subsequently, some works [50,53] explore knowledge decoupling and perform linear parameter fusion with re- parameterization . However, this is an intra-module fusion strategy, which is restricted to certain operations. As the last category, we propose Endpoints Weight Fusion (EWF) in the form of parameter addition between the old and new model with a dynamic factor, which requires no further training and re-parameterization, and maintains a constant model size as more tasks are encountered. In this work, we adapt weight fusion to CISS and pro- pose the EWF strategy, which aims at utilizing weight fu- sion to find a new balance between old and new knowledge. During incremental training, we choose a starting point and an ending point model of the current task training trajectory. The starting point represents the old knowledge, while the ending point represents the new knowledge. After learn- ing the current task, a dynamic weight fusion is proposed for efficient knowledge integration. We aggregate them by taking the weighted average of the corresponding parame- ters of the two models. Nevertheless, the training procedure without restraints on the model would increase the param- eter distance between the start and end points, limiting the performance improvement brought by the EWF strategy. To overcome this shortcoming, we further enhance the EWF strategy with a knowledge distillation scheme [16, 17, 50], which can largely increase the similarity of the models at the two points and boost the efficiency of EWF. To summarize, the main contributions of this paper are: • We propose an Endpoints Weight Fusion strategy, which has no cost of further training and keeps the model size the same. It can effectively find a new balance between old and new categories and alleviate catastrophic forgetting. • Our method can be easily integrated with several state- of-the-art methods. In several CISS scenarios of long sequences, it can boost the baseline performance bymore than 20%. • We conduct experiments on various CISS scenarios, which demonstrate that our method achieves the state- of-the-art performance on both PASCAL VOC and ADE20K.
Wang_Feature_Alignment_and_Uniformity_for_Test_Time_Adaptation_CVPR_2023	Abstract Test time adaptation (TTA) aims to adapt deep neural networks when receiving out of distribution test domain samples. In this setting, the model can only access online unlabeled test samples and pre-trained models on the train- ing domains. We first address TTA as a feature revision problem due to the domain gap between source domains and target domains. After that, we follow the two measure- ments alignment anduniformity to discuss the test time fea- ture revision. For test time feature uniformity, we propose atest time self-distillation strategy to guarantee the consis- tency of uniformity between representations of the current batch and all the previous batches. For test time feature alignment, we propose a memorized spatial local cluster- ingstrategy to align the representations among the neigh- borhood samples for the upcoming batch. To deal with the common noisy label problem, we propound the entropy and consistency filters to select and drop the possible noisy la- bels. To prove the scalability and efficacy of our method, we conduct experiments on four domain generalization bench- marks and four medical image segmentation tasks with var- ious backbones. Experiment results show that our method not only improves baseline stably but also outperforms ex- isting state-of-the-art test time adaptation methods.	1. Introduction Deep learning has achieved great success in computer vision tasks when training and test data are sampled from the same distribution [ 17,26,37]. However, in real-world applications, performance degradation usually occurs when training (source) data and test (target) data are collected from different distributions, i.e.domain shift . In practice, test samples may encounter different types of variations or corruptions. Deep learning models will be sensitive to these variations or corruptions, which can cause perfor- mance degradation. To tackle this challenging but practical problem, variousworks have been proposed to adapt the model at testtime [ 5, 8,21,36,43,55,64,65,74]. Test time training (TTT) [ 36,55] adapts model using self-supervised tasks, such as rotation classification [ 15] during both training and test phases. This paradigm relies on additional model modifications in both training and test phases, which is not feasible and scalable in the real world. Similar to TTT, test time adaptation [ 64] (TTA) also adapts the model i.e., updates the parameters in the test phase. But TTA does not require any specific modifications in training and requires only the pre-trained source model and unlabeled target data during the test phase, which is more practical and generalizable. In TTA, the model could be adapted with only online unlabeled data. Hence, the model trained on source data is incompatible with the target data due to the possible domain gap between source data and target data. To deal with the above-mentioned problem, we address TTA as a representation revision problem in this paper. In the test phase of TTA, the accessed model has already learned the feature representations specialized to source do- mains and may generate inaccurate representations for the target domain due to the large domain gap. It is necessary to rectify the feature representations for the target domain. To achieve better representations for the target domain, we utilize the commonly used measurements for representation quality that can be summarized to feature alignment and uniformity [66,71]. Alignment refers that the similar im- ages should have the similar representations, while unifor- mity means that images of different classes should be dis- tributed as uniform as possible in the latent space. Hence, we propose to address the TTA problem from the above- mentioned properties. Most of the previous works about TTA can be in- ducted from the proposed representation revision perspec- tive. Some methods adapt the source model by conduct- ing a feature alignment process, such as feature matching [25,36] and predictions adjustment [ 5]. One of the repre- sentative methods is LAME [ 5], which encourages neigh- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20050 borhood samples in the feature space to have similar pre- dictions using Laplacian adjusted maximum-likelihood es- timation. Moreover, other methods aim to make target feature more uniform in feature space, including entropy minimization [ 43,64,74], prototype adjustment [ 21], infor- mation maximization [ 34] and batch normalization statis- tics alignment [ 33,42,51]. One representative method is T3A [ 21], which adjusts prototypes (class-wise centroids) to have a more uniform representation by building a sup- port set. However, none of the method address the TTA problem from the representation alignment and uniformity simultaneously. In this paper, we identify this limitation and propose a novel method that rectifies feature representation from both two properties. We formulate the two properties in TTA as test time feature uniformity andtest time feature alignment . Test Time Feature Uniformity. Following the feature uniformity perspective, we hope that representations of test images from different classes should be distributed as uni- form as possible. However, only limited test samples can be accessed in an online manner in the TTA setting. To better deal with all of the samples in the target do- main, we propose to introduce historical temporal informa- tion for every arriving test sample. A memory bank is built to store feature representations and logits for all arriving samples to maintain the useful information from the previ- ous data. We then calculate the pseudo-prototypes for every class by using the logits and features in the memory bank. After that, to guarantee the uniformity for the current batch of samples, the prediction distribution of prototype-based classification and model prediction (outputs of linear clas- sifier) should be similar, i.e., the feature distribution of the current images of one class should be consistent with the feature distribution of all the previous images of the same class. This can reduce the bias of misclassified outlier sam- ples to form a more uniform latent space. Motivated by this, we minimize the distance between the outputs of linear and prototype-based classifiers. This pattern is similar to self-distillation [ 23,73] that transfers knowledge between different layers of the same network architecture. However, unlike typical self-distillation, our method does not require any ground truth supervision. We refer to this method as Test Time Self-Distillation (TSD). Test Time Feature Alignment. Feature alignment en- courages the images from the same class to have similar fea- ture representations in the latent space. As for TTA, pseudo labels generated by the source model may be noisy due to the domain gap. Thus, instead of aligning all the positive pairs, we propose a K-nearest feature alignment to encour- age features from the same class to be closed or features from different classes to be far away from each other. This can reduce the negative impact imposed by the noisy la- bels and maintain the alignment of images with the samesemantics. Specifically, we retrieve K-nearest features in the memory bank for the upcoming images and add consis- tency regularization between the representations and logits of the images. We refer to this as Memorized Spatial Local Clustering (MSLC). The ablation of hyperparameter Kis shown in Table 5and Fig. 3. Entropy Filter and Consistency Filter. During the adaption, we use stored pseudo features and logits to com- pute the pseudo-prototypes. However, the noisy label prob- lem cannot be completely alleviated despite the proposed efforts. To further reduce the impact, we adopt both entropy andconsistency filters to filter noisy labels to boost perfor- mance. As for the entropy filter, we filter noisy features with high entropy when we compute prototypes because unreli- able samples usually produce high entropy. In addition, the predictions of prototype-based and lin- ear classifiers of the network for reliable samples should be consistent ideally. We use this property to filter unreliable samples and back-propagate the gradient using only reliable samples. We refer to this filter as the consistency filter. The ablation study on the two proposed filters is presented in Table 5. Finally, we demonstrate the effectiveness of our pro- posed approach on commonly used domain generalization benchmarks, including PACS [ 30], VLCS [ 58], OfficeHome [62] and DomainNet [ 47]. Furthermore, to prove the effi- cacy of our method, we conduct more experiments on four cross-domain medical image segmentation benchmarks, in- cluding prostate segmentation [ 1,35], cardiac structure segmentation [ 4,76,77] and optic cup/disc segmentation [13,45,53]. Our method achieves the state-of-the-art per- formance on the above benchmarks. We summarize our contributions as follows: • We propose a new perspective for test time adapta- tion from the view of feature alignment and unifor- mity. The proposed test time feature uniformity en- courages the representations of the current batch of samples along with the uniformity of all the previous samples. The test time feature alignment manipulates the representation of the test sample according to its neighbors in the latent space to align the representa- tions based on the pseudo label. • Specifically, to meet the online setting and noisy la- bel problem in TTA, we propose two complementary strategies: unsupervised self-distillation for test time feature uniformity and memorized spatial local cluster- ing for test time feature alignment. We also propound the entropy filter and consistency filter to further miti- gate the effect of the noisy labels. • The experiments demonstrate that our proposed method outperforms existing test time adaptation ap- 20051 proaches on both the domain generalization bench- marks and medical image segmentation benchmarks.
Wasim_Vita-CLIP_Video_and_Text_Adaptive_CLIP_via_Multimodal_Prompting_CVPR_2023	Abstract Adopting contrastive image-text pretrained models like CLIP towards video classification has gained attention due to its cost-effectiveness and competitive performance. How- ever, recent works in this area face a trade-off. Fine- tuning the pretrained model to achieve strong supervised performance results in low zero-shot generalization. Sim- ilarly, freezing the backbone to retain zero-shot capabil- ity causes significant drop in supervised accuracy. Be- cause of this, recent works in literature typically train sep- arate models for supervised and zero-shot action recog- nition. In this work, we propose a multimodal prompt learning scheme that works to balance the supervised and zero-shot performance under a single unified train- ing. Our prompting approach on the vision side caters for three aspects: 1) Global video-level prompts to model the data distribution; 2) Local frame-level prompts to provide per-frame discriminative conditioning; and 3) a summary prompt to extract a condensed video representation. Ad- ditionally, we define a prompting scheme on the text side to augment the textual context. Through this prompting scheme, we can achieve state-of-the-art zero-shot perfor- mance on Kinetics-600, HMDB51 and UCF101 while re- maining competitive in the supervised setting. By keep- ing the pretrained backbone frozen, we optimize a much lower number of parameters and retain the existing gen- eral representation which helps achieve the strong zero- shot performance. Our codes/models will be released at https://github.com/TalalWasim/Vita-CLIP ..	1. Introduction In the image classification domain, multimodal image- text pretrained models such as CLIP [58], ALIGN [31] and Florence [75] have shown the capability of learning gener- alized representations. These models, trained on large-scale language-image pairs in a contrastive manner, have remark- able zero-shot capabilities and transfer well to a variety ofdownstream tasks. However, training a similar model for the task of video recognition is not feasible both in terms of gathering large-scale video-text pairs, which can suffer from alignment problems [30], and is also exponentially more computationally expensive due to multiple frames be- ing processed per video. Therefore, there has been a recent push in the research community to effectively adopt the pre- trained image-text models for the task of video recognition, while maintaining their zero-shot capabilities. In this re- gard, existing methods can be divided into two categories. Some take inspiration from recent prompt learning methods [25, 32, 77, 81, 82] and propose a prompt learning scheme either on the text [36] or vision [55, 70] side, along with additional transformer layers for improved temporal learn- ing. Others prefer an end-to-end CLIP finetuning scheme for video tasks [51, 55, 70]. However, the problem with these methods is that they either fail to effectively leverage learning on both the text and vision sides [36, 55] or end up losing the zero-shot generalization of CLIP by finetun- ing the vision decoder [47] or the backbone [51, 55, 70]. In summary, the existing approaches can steer the model either towards good zero-shot generalization orbetter supervised learning on video tasks. Since real-world tasks require both supervised and zero-shot capabilities, our work investigates the following question: Can we develop a unified model for videos that performs well for both supervised learning and zero-shot generalization tasks? In pursuit of the aforementioned question, we propose a multimodal prompting-based Video and text adaptive CLIP. To effectively adapt the pretrained image-text CLIP model to videos, we consider two important aspects. Firstly, one needs to preserve the generalization capabilities of the orig- inal pretrained CLIP backbone and secondly, it must be able to effectively adapt to the video domain. In this regard, we propose to keep the entire backbone frozen and learn addi- tional lightweight modules to adapt the model for videos. On this point, for the vision side, we aim to explicitly ex- ploit the temporal information in videos which is lacking in the frozen image model. Our approach models video infor- mation at three levels: first via global video-level prompts This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23034 (a) Proposed Prompting Scheme76 78 80 82 8440424446485052 Vita-CLIP B/16 (OURS) (frozen backbone) ActionClip B/16 [70] (CORR’21) (finetuned backbone)XCLIP B/16 [55] (ECCV’22) (finetuned backbone)A5[36] (ECCV’22) (frozen backbone) Top-1 K400 SupervisedTop-1 HMDB51 Zeroshot 76 78 80 82 84556065707580 Vita-CLIP B/16 (OURS) (frozen backbone) ActionClip B/16 [70] (CORR’21) (finetuned backbone)XCLIP B/16 [55] (ECCV’22) (finetuned backbone)A5[6](ECCV’22) (frozen backbone) Top-1 K400 SupervisedTop-1 UCF101 Zeroshot (b) Zero-shot accuracy (HMDB51, UCF101) vs supervised accuracy (Kinetics-400).Figure 1. An overview of the proposed prompting scheme (left) alongside the trade- off which we attempt to balance between su- pervised and zero-shot performance ( right ). (a)Our prompting approach adds learnable pa- rameters to learn visual and temporal infor- mation in videos at three levels: a summary prompt to learn a condensed representation of the video, video-level prompts to model global distribution shifts needed to adapt to video do- main and frame-level prompts to enrich local discriminative information in each frame. On the text side, we learn prompts to adapt the language representations for videos. (b)The trade-off plots showing zero-shot vs. super- vised performance comparison for ours and re- cent CLIP-based video approaches. Note that existing SoTA [55] trains two separate models for zero-shot and supervised settings while our method offers a unified model with the same training for both settings. that learn the overall distribution characteristics of video data e.g., motion and dynamics; secondly, inspired by [53], local frame-level prompts which model per frame discrimi- native information by directly conditioning on classification tokens of all frames; and thirdly by a summary prompt that distills the entire video sequence response in a single con- cise summary vector. Additionally, to better model the textual context we pro- pose to use a learnable context on the text encoder. The rea- son why this is particularly important is that the textual in- formation is quite limited in the available video datasets. In- stead of having per-sample text descriptions, we are limited to using class labels as text descriptions. Inspired by [82], we propose a prompt learning method on the text side to better model the textual context and to augment the video class label descriptions. An overview of our method with the trade-off it seeks to balance is presented in Fig. 1. The main contributions of this work are as follows: • We propose a multimodal prompting approach Vita-CLIP for videos that learns video and text-specific context vec- tors to efficiently adapt the image-text pretrained CLIP model to video recognition tasks. • On the vision side, we explicitly model the temporal in- formation and the video data distribution. Our prompt learning method aggregates the discriminative informa- tion from each frame in a clip with every other frame, while also providing per-layer learning capacity to better capture the data distribution. On the language side, our approach learns complimentary semantic context to bet- ter adapt the language representations. • We evaluate our approach on supervised as well as gener-alization tasks and demonstrate a sound balance between both aspects using a single unified model. Specifically, on zero-shot tasks, we obtain 4.0%, 3.0% and 2.2% gains over the recent SoTA X-CLIP [55] on HMDB-51, UCF- 101, and Kinetics-600 datasets respectively.
Wang_ALTO_Alternating_Latent_Topologies_for_Implicit_3D_Reconstruction_CVPR_2023	Abstract This work introduces alternating latent topologies (ALTO) for high-fidelity reconstruction of implicit 3D sur- faces from noisy point clouds. Previous work identifies that the spatial arrangement of latent encodings is important to recover detail. One school of thought is to encode a la- tent vector for each point (point latents). Another school of thought is to project point latents into a grid (grid la- tents) which could be a voxel grid or triplane grid. Each school of thought has tradeoffs. Grid latents are coarse and lose high-frequency detail. In contrast, point latents preserve detail. However, point latents are more difficult to decode into a surface, and quality and runtime suffer. In this paper, we propose ALTO to sequentially alternate between geometric representations, before converging to an easy-to-decode latent. We find that this preserves spa- tial expressiveness and makes decoding lightweight. We validate ALTO on implicit 3D recovery and observe not only a performance improvement over the state-of-the-art, but a runtime improvement of 3-10×. Project website at https://visual.ee.ucla.edu/alto.htm/ . *Equal contribution.	1. Introduction Reconstructing surfaces from noisy point clouds is an active problem in 3D computer vision. Today, conditional neural fields offer a promising way to learn surfaces from noisy point clouds. Alternatives like voxel regression or mesh estimation are limited by cubic complexity and the requirement of a mesh template, respectively. Recent work has successfully used conditional neural fields to recon- struct 3D surfaces as an occupancy function. A conditional neural field takes as input a query coordinate and conditions this on a latent representation, e.g., feature grids. The spa- tial expressiveness of the latent representation impacts the overall surface reconstruction quality. To achieve spatial expression, a neural field is condi- tioned on a latent space of features (latents) from the con- ditional input. In 3D surface reconstruction the input point cloud is transformed into latents arranged in some topolog- ical structure. Point latents occur when each point in the input point cloud is assigned a latent vector [4]. Triplane latents are formed when point latents are projected into a 3-axis grid [41, 52]. The triplane latent is not as spatially expressive as freeform points, but the lower spatial com- plexity makes it easier to decode. Voxel latents are another type of grid latent where latents are arranged in a feature This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 259 Figure 2. An overview of our method. Given input surface points, we obtain an implicit occupancy field with iterative alternation between features in the forms of points and 2D or 3D grids (Sec. 3.2). Then we decode the occupancy values for query points with a learned attention-based interpolation from neighboring grids (Sec. 3.3). volume [52, 66]. To reconstruct detailed surfaces, recent state-of-the-art methods try to preserve point latents as long as possi- ble. Because point latents are spatially expressive, methods based on point latents are considered state-of-the-art for de- tailed surface reconstruction [4, 18]. However, using point latents in this way has some tradeoffs. It is difficult to cor- relate a query with the unstructured topology of a point- based latent space, placing a burden on the decoder. Results from POCO [4] are shown in Fig. 1where runtime and high- quality detail like thin lampposts remain out of reach. In this paper, we seek to blend the upside of different latent topologies, while minimizing downside. We present an alternating latent topology (ALTO) method. In contrast to previous work, our method does not stay with either point [4] or grid latents [52], but instead alternates back and forth between point and grid latents before converging to a final grid for ease-of-decoding. Our method is general. We can plug-in the ALTO com- ponent to existing grid-based conditional models [10, 52] to boost detail recovery. While we have shown that our method can generate occupancy fields, we expect gain of high-fidelity details for other neural fields, such as seman- tic or affordance fields [32, 70], where similar conditional techniques can be adopted. We summarize our contributions as follows: • We introduce an iterative technique to blend the strengths of different latent topologies for high-fidelity conditional neural fields generation. • We propose an attention-based decoder that replaces naive linear interpolation of feature-grids or computa- tionally expensive point-wise attention while keeping compute burden in check. • We demonstrate performance and runtime improve- ments over the highest-quality previous method [4], as well as performance improvements over all other base- lines.
Wang_Neural_Residual_Radiance_Fields_for_Streamably_Free-Viewpoint_Videos_CVPR_2023	Abstract The success of the Neural Radiance Fields (NeRFs) for modeling and free-view rendering static objects has in-spired numerous attempts on dynamic scenes. Current tech-niques that utilize neural rendering for facilitating free-view videos (FVVs) are restricted to either ofﬂine render-ing or are capable of processing only brief sequences with minimal motion. In this paper , we present a novel tech- nique, Residual Radiance Field or ReRF , as a highly com-pact neural representation to achieve real-time FVV ren- dering on long-duration dynamic scenes. ReRF explicitlymodels the residual information between adjacent times- tamps in the spatial-temporal feature space, with a global coordinate-based tiny MLP as the feature decoder . Specif- ically, ReRF employs a compact motion grid along with aresidual feature grid to exploit inter-frame feature similar- ities. We show such a strategy can handle large motions without sacriﬁcing quality. We further present a sequentialtraining scheme to maintain the smoothness and the spar- sity of the motion/residual grids. Based on ReRF , we design †The corresponding authors are Minye Wu (minye.wu@kuleuven.be) and Lan Xu (xulan1@shanghaitech.edu.cn).a special FVV codec that achieves three orders of magni- tudes compression rate and provides a companion ReRF player to support online streaming of long-duration FVVs of dynamic scenes. Extensive experiments demonstrate theeffectiveness of ReRF for compactly representing dynamicradiance ﬁelds, enabling an unprecedented free-viewpoint viewing experience in speed and quality.	1. Introduction Photo-realistic free-viewpoint videos (FVVs) of dy- namic scenes, in particular, human performances, reduce the gap between the performer and the viewer. But the goal of producing and viewing FVVs as simple as clicking andviewing regular 2D videos on streaming platforms remains far-reaching. The challenges range from data processing and compression to streaming and rendering. Geometry-based solutions reconstruct dynamic 3D meshes or points [ 14,16], whereas image-based ones inter- polate novel views on densely transmitted footages [ 6,83]. Both techniques rely on high-quality reconstructions thatare often vulnerable to occlusions and textureless regions. Recent neural advances [ 44,61] bring an alternative route This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 76 that bypasses explicit geometric reconstruction. The sem- inal work of the Neural Radiance Field (NeRF) [ 44] com- pactly represents a static scene in a coordinate-based multi-layer perceptron (MLP) to conduct volume rendering at photo-realism. The MLP can be viewed as an implicit fea- ture decoder from a spatially continuous feature space to the radiance output with RGB and density. However, us-ing even a moderately deep MLP can be too expensive forreal-time rendering. V arious extensions have hence focusedon “sculpting” the feature space using smart representationsto strike an intricate balance between computational speedand accuracy. Latest examples include explicit feature vol-umes [ 21,57,77], multi-scale hashing [ 45], codebook [ 59], tri-planes [ 8], tensors [ 11,60], etc. Although effective, by far nearly all methods are tailored to handle static scenes. In contrast, streaming dynamic radi-ance ﬁelds require using a global coordinate-based MLP to decode features from a spatial-temporally continuous fea-ture space into radiance outputs. A na ¨ıve per-frame solu- tion would be to apply static methods [ 45,60] on a series of independent spatial feature spaces. Such schemes discard important temporal coherency, yielding low quality and in- efﬁciency for long sequences. Recent methods attempt to maintain a canonical feature space to reproduce features in each live frame by temporally warping them back into thecanonical space. V arious schemes to compensate for tem- poral motions have been proposed by employing implicitmatching [ 18,38,48,49,62] or data-driven priors such as depth [ 73], Fourier features [ 67], optical ﬂow [ 17,37], or skeletal/facial motion priors [ 28,50,69,82]. However, heavy reliance on the global canonical space makes them fragileto large motions or topology changes. The training over- head also signiﬁcantly increases according to the sequencelength. Recent work [ 34] sets out to explore feature redun- dancy between adjacent frames but it falls short of main-taining a coherent spatial-temporal feature space. In this paper, we present a novel neural modeling tech- nique that we call the Residual Radiance Field or ReRF as a highly compact representation of dynamic scenes, enablinghigh-quality FVV streaming and rendering (Fig. 1). ReRF explicitly models the residual of the radiance ﬁeld betweenadjacent timestamps in the spatial-temporal feature space.Speciﬁcally, we employ a global tiny MLP to approximateradiance output of the dynamic scene in a sequential man-ner. To maintain high efﬁciency in training and inference,ReRF models the feature space using an explicit grid rep-resentation analogous to [ 57]. However, ReRF only per- forms the training on the ﬁrst key frame to obtain an MLPdecoder for the whole sequence and at the same time it uses the resulting grid volume as the initial feature volume. For each subsequent frame, ReRF uses a compact motion grid and a residual feature grid: the low-resolution motion gridrepresents the position offset from the current frame to theprevious whereas a sparse residual grid is used to compen- sate for errors and newly observed regions. A major beneﬁt of such a design is that ReRF fully exploits feature similar-ities between adjacent frames where the complete featuregrid of the current frame can be simply obtained from thetwo while avoiding the use of a global canonical space. Inaddition, both motion and residual grids are amenable forcompression, especially for long-duration dynamic scenes. We present a two-stage scheme to efﬁciently obtain the ReRF from RGB videos via sequential training. In particu-lar, we introduce a novel motion pooling strategy to main- tain the smoothness and compactness of the inter-frame mo- tion grid along with sparsity regularizers to improve thecompactness of ReRF. To make ReRF practical for users, we further design a ReRF-based codec that follows the traditional keyframe-based strategy, achieving three ordersof magnitudes compression rate compared to per-frame-based neural representations [ 57]. Finally, we demonstrate a companion ReRF player suitable for conducting onlinestreaming of long-duration FVVs of dynamic scenes. WithReRF, a user, for the ﬁrst time, can pause, play, fast for- ward/backward, and seek on dynamic radiance ﬁelds as if viewing 2D videos, resulting in an unprecedented high-quality free-viewpoint viewing experience (see Fig. 2 ). To summarize, our contributions include: • We introduce Residual Radiance Field (ReRF), a novel neural representation, to support streamable free-viewpoint viewing of dynamic radiance ﬁelds. • We present tailored motion and residual grids to sup- port sequential training and at the same time eliminate the need for using a global canonical space notorious for large motions. We further introduce a number oftraining strategies to achieve a high compression rate while maintaining high rendering quality. • We develop a ReRF-based codec and a companion FVV player to stream dynamic radiance ﬁelds of long sequences, with broad control functions.
Wang_YOLOv7_Trainable_Bag-of-Freebies_Sets_New_State-of-the-Art_for_Real-Time_Object_Detectors_CVPR_2023	Abstract Real-time object detection is one of the most important research topics in computer vision. As new approaches re- garding architecture optimization and training optimization are continually being developed, we have found two re- search topics that have spawned when dealing with these latest state-of-the-art methods. To address the topics, we propose a trainable bag-of-freebies oriented solution. We combine the ﬂexible and efﬁcient training tools with the proposed architecture and the compound scaling method. YOLOv7 surpasses all known object detectors in both speed and accuracy in the range from 5 FPS to 120 FPS and has the highest accuracy 56.8% AP among all known real- time object detectors with 30 FPS or higher on GPU V100. Source code is released in https://github.com/ WongKinYiu/yolov7 .	1. Introduction Real-time object detection is a very important topic in computer vision, as it is often a necessary component in computer vision systems. For example, multi-object track- ing [90, 91], autonomous driving [17, 39], robotics [34, 55], medical image analysis [33, 44], etc. The computing de- vices that execute real-time object detection is usually some mobile CPUs or GPUs, as well as various neural process- ing units (NPUs). For example, the Apple neural engine (Apple), the neural compute stick (Intel), Jetson AI edge devices (Nvidia), the edge TPU (Google), the neural pro- cessing engine (Qualcomm), the AI processing unit (Medi- aTek), and the AI SoCs (Kneron), are all NPUs. Some of edge devices focus on speeding up different operations such as vanilla convolution, depth-wise convolution, or MLP op- erations. In this paper, the real-time object detector we pro- posed mainly hopes that it can support both mobile GPU and GPU devices from the edge to the cloud. In recent years, the real-time object detector is still devel- oped for different edge devices. For example, the develop- Figure 1. Comparison with other real-time object detectors, our proposed methods achieve state-of-the-arts performance. ment of MCUNet [46,47] and NanoDet [51] focused on pro- ducing low-power single-chip and improving the inference speed on edge CPU. As for methods such as YOLOX [20] and YOLOR [79], they focus on improving the inference speed of various GPUs. More recently, the development of real-time object detector has focused on the design of ef- ﬁcient architecture. As for real-time object detectors that can be used on CPU [51, 81, 82, 86], their design is mostly based on MobileNet [26, 27, 63], ShufﬂeNet [52, 89], or GhostNet [24]. Another real-time object detectors are de- veloped for GPU [20, 79, 94], they mostly use ResNet [25], DarkNet [60], or DLA [85], and then use the CSPNet [77] strategy to optimize the architecture. The development di- rection of the proposed methods in this paper are different from that of the current real-time object detectors. In ad- dition to architecture optimization, our proposed methods will focus on the optimization of the training process. Our focus will be on some optimized modules and optimization methods which may strengthen the training cost for improv- ing the accuracy of object detection, but without increasing the inference cost. We call these modules and optimization methods trainable bag-of-freebies. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7464 Recently, model re-parameterization [11, 12, 28] and dy- namic label assignment [16, 19, 40] have become important topics in network training and object detection. Mainly af- ter the above new concepts are proposed, the training of object detector evolves many new issues. In this paper, we will present some of the new issues we have discovered and devise effective methods to address them. For model re- parameterization, we analyze the model re-parameterization strategies applicable to layers in different networks with the concept of gradient propagation path, and propose planned re-parameterization model. In addition, when we discover that with dynamic label assignment technology, the train- ing of model with multiple output layers will generate new issues. That is: “How to assign dynamic targets for the out- puts of different branches?” For this problem, we propose a new label assignment method called coarse-to-ﬁne lead guided label assignment. The contributions of this paper are summarized as fol- lows: (1) we design several trainable bag-of-freebies meth- ods, so that real-time object detection can greatly improve the detection accuracy without increasing the inference cost; (2) for the evolution of object detection methods, we found two new issues, namely how re-parameterization module replaces original module, and how dynamic label assignment strategy deals with assignment to different out- put layers. In addition, we also propose methods to address the difﬁculties arising from these issues; (3) we propose “extend” and “compound scaling” methods for the real-time object detector that can effectively utilize parameters and computation; and (4) the method we proposed can effec- tively reduce large amount of parameters and computation of state-of-the-art real-time object detector, and has faster inference speed and higher detection accuracy.
Wu_Referring_Multi-Object_Tracking_CVPR_2023	Abstract Existing referring understanding tasks tend to involve the detection of a single text-referred object. In this paper, we propose a new and general referring understanding task, termed referring multi-object tracking (RMOT). Its core idea is to employ a language expression as a semantic cue to guide the prediction of multi-object tracking. To the best of our knowledge, it is the ﬁrst work to achieve an arbitrary number of referent object predictions in videos. To push forward RMOT, we construct one benchmark with scalable expressions based on KITTI, named Refer-KITTI. Speciﬁ- cally, it provides 18 videos with 818 expressions, and each expression in a video is annotated with an average of 10.7 objects. Further, we develop a transformer-based architec- ture TransRMOT to tackle the new task in an online manner, which achieves impressive detection performance and out- performs other counterparts. The Refer-KITTI dataset and the code are released at https://referringmot.github.io .	1. Introduction Recently, referring understanding [ 5,17,33,55], inte- grating natural language processing into scene perception, has raised great attention in computer vision community. It aims to localize regions of interest in images or videos under the instruction of human language, which has many applications, such as video editing and autonomous driving. For referring understanding, several signiﬁcant benchmarks have been published. Flickr30k [ 53], ReferIt [ 15], and Re- fCOCO/+/g [ 55] have greatly encouraged the development of image-based referring tasks. More datasets ( e.g., Lin- gual OTB99 [ 19], Cityscapes-Ref [ 42], Talk2Car [ 5], Refer- DA VIS 17[17], and Refer-Youtube-VOS [ 38]) are further ∗Equal contribution. †Corresponding author: Jianbing Shen . This work was supported in part by the FDCT grant SKL-IOTSC(UM)-2021- 2023, the Grant MYRG-CRG2022-00013-IOTSC-ICI, and the Start-up Research Grant (SRG) of University of Macau (SRG2022-00023-IOTSC). ‡The work is done during the internship at MEGVII Technology. (a) Query : the cars in the right (b) Query : the cars which are turning Figure 1. Representative examples from RMOT . The expression query can refer to multiple objects of interest (a), and captures the short-term status with accurate labels (b). proposed to cover the application in videos. Despite these advanced progress, previous benchmarks have two typical limitations. First , each expression tends to correspond to only one target. However, many objects have the same semantics in an open world, i.e., one sin- gle expression could refer to multiple objects. From this side, existing datasets lack ﬂexible simulation on the multi- object scenarios, causing referring understanding tasks far from satisfactory. Second , the given expression may only describe part of frames for the video referring task, mak- ing the correspondence inaccurate. For example, given the expression ‘the car which is turning’, we have to predict the overall trajectory even if the car has ﬁnished the turn- ing action. Obviously, a single expression cannot cover all short-term status of one target. Overall, existing datasets fail to provide an accurate evaluation under the situations of multiple referent targets and temporal status variances. To address these problems, we propose a novel video un- derstanding task guided by the language description, named referring multi-object tracking (RMOT). Given a language expression as a reference, it targets to ground all semanti- cally matched objects in a video. Unlike previous tasks, our proposed RMOT is much closer to the real environment, as each expression can involve multiple objects. For in- stance, the expression query ‘the cars in the right’ corre- sponds to one object at the 20thframe but two objects at the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 14633 The red cars The pedestrian The persons in the right The cars which are turning The black cars which are moving The cars which are slower than oursThe parking cars The cars in the counter direction of ours The cars in leftFigure 2. More examples of Refer-KITTI . It provides high-diversity scenes and high-quality annotations referred to by expressions. 40thframe (see Fig. 1(a)). The phenomenon indicates that RMOT focuses more on ﬁnding the matched targets so that the referent number can be ﬂexibly changed. In addition, the temporal status variances are also considered in RMOT. As shown in Fig. 1(b), the given example shows the cars can be detected only when they start the turning action, and the tracking will be ended if they ﬁnish the activity. To speed up the development of RMOT, we construct a new benchmark, i.e., Refer-KITTI, concerning the trafﬁc scenes. It is developed from the public KITTI [ 9] dataset. Compared to existing referring understanding datasets, it has three distinguishing characteristics: i)High ﬂexibility with referent objects. The number of objects described by each expression range from 0 to 105, with 10.7 on average. ii)High temporal dynamics. The temporal status of targets covers a longer time with more frames (varying in 0 ∼400 frames), and the temporal variance of targets is accurately captured using our labeling tool. iii)Low labeling cost with identiﬁcation spread. We provide an effortless tool to anno- tate a target tracklet using only two clicks. Although RMOT has a more ﬂexible referring setting, it brings additional challenges: multi-object prediction and cross-frame association. Towards this end, we propose an end-to-end differentiable framework for RMOT. Our model builds upon the recent DETR framework [ 3], enhanced by powerful cross-modal reasoning and cross-frame conjunc- tion. It has an encoder-decoder architecture. Speciﬁcally, we design an early-fusion module in the encoder to densely integrate visual and linguistic features, followed by a stack of deformable attention layers for further reﬁning the cross- modal representations. In the decoder, query-based embed- dings interact with the cross-modal features to predict ref- erent boxes. To track multi-objects, similar to MOTR [ 57], we decouple the object queries into track query for tracking objects of previous frames and detect query for predicting the bounding boxes of new-born objects. In summary, our contributions are three-fold. First , we propose a new task for referring multi-objects, called re-Dataset Video ImagesInstances per-expressionTemporal ratio per-expression RefCOCO [ 55] - 26,711 1 1 RefCOCO+ [ 55] - 19,992 1 1 RefCOCOg [ 55] - 26,711 1 1 Talk2Car [ 5] ✓ 9,217 1 - VID-Sentence [ 4]✓ 59,238 1 1 Refer-DA VIS 17[17]✓ 4,219 1 1 Refer-YV [ 38] ✓ 93,869 1 1 Refer-KITTI ✓ 6,650 10.7 0.49 Table 1. Comparison of Refer-KITTI with existing datasets. Refer-YV is short for Refer-Youtube-VOS. The temporal ratio rep- resents the average ratio of referent frames covering the entire video sequence. ‘-’ means unavailable. ferring multi-object tracking (RMOT). It tackles limitations in the existing referring understanding tasks and provides multi-object and temporally status-variant circumstances. Second , we formulate a new benchmark, Refer-KITTI, to help the community to explore this new ﬁeld in depth. As far as we know, it is the ﬁrst dataset specializing in an ar- bitrary number of object predictions. Third , we propose an end-to-end framework built upon Transformer, termed as TransRMOT. With powerful cross-modal learning, it pro- vides impressive RMOT performance on Refer-KITTI com- pared to hand-crafted RMOT methods.
Wen_BundleSDF_Neural_6-DoF_Tracking_and_3D_Reconstruction_of_Unknown_Objects_CVPR_2023	Abstract We present a near real-time (10Hz) method for 6-DoF tracking of an unknown object from a monocular RGBD video sequence, while simultaneously performing neural 3D reconstruction of the object. Our method works for arbi- trary rigid objects, even when visual texture is largely ab- sent. The object is assumed to be segmented in the first frame only. No additional information is required, and no assumption is made about the interaction agent. Key to our method is a Neural Object Field that is learned concur- rently with a pose graph optimization process in order to robustly accumulate information into a consistent 3D rep- resentation capturing both geometry and appearance. A dy- namic pool of posed memory frames is automatically main- tained to facilitate communication between these threads. Our approach handles challenging sequences with large pose changes, partial and full occlusion, untextured sur- faces, and specular highlights. We show results on HO3D, YCBInEOAT, and BEHAVE datasets, demonstrating that our method significantly outperforms existing approaches. Project page: https://bundlesdf.github.io/	1. Introduction Two fundamental (and closely related) problems in com- puter vision are 6-DoF (“degree of freedom”) pose tracking and 3D reconstruction of an unknown object from a monoc- ular RGBD video. Solving these problems will unlock a wide range of applications in areas such as augmented reality [34], robotic manipulation [22, 70], learning-from- demonstration [71], and sim-to-real transfer [1, 15].Prior efforts often consider these two problems sepa- rately. For example, neural scene representations have achieved great success in creating high quality 3D object models from real data [3, 40, 44, 59, 68, 81]. These ap- proaches, however, assume known camera poses and/or ground-truth object masks. Furthermore, capturing a static object by a dynamically moving camera prevents full 3D reconstruction ( e.g., the bottom of the object is never seen if resting on a table). On the other hand, instance-level 6-DoF object pose estimation and tracking methods of- ten require a textured 3D model of the test object before- hand [24, 28, 66, 72, 73] for pre-training and/or online tem- plate matching. While category-level methods enable gen- eralization to new object instances within the same cate- gory [7,27,62,67,74], they struggle with out-of-distribution object instances and unseen object categories. To overcome these limitations, in this paper we pro- pose to solve these two problems jointly. Our method as- sumes that the object is rigid, and it requires a 2D object mask in the first frame of the video. Apart from these two requirements, the object can be moved freely through- out the video, even undergoing severe occlusion. Our ap- proach is similar in spirit to prior work in object-level SLAM [35, 36, 50–52, 64, 85], but we relax many common assumptions, allowing us to handle occlusion, specularity, lack of visual texture and geometric cues, and abrupt object motion. Key to our method is an online pose graph opti- mization process, a concurrent Neural Object Field to re- construct the 3D shape and appearance, and a memory pool to facilitate communication between the two processes. The This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 606 robustness of our method is highlighted in Fig. 1. Our contributions can be summarized as follows: •A novel method for causal 6-DoF pose tracking and 3D reconstruction of a novel unknown dynamic object. This method leverages a novel co-design of concurrent track- ing and neural reconstruction processes that run online in near real-time while largely reducing tracking drift. •We introduce a hybrid SDF representation to deal with uncertain free space caused by the unique challenges in a dynamic object-centric setting, such as noisy segmenta- tion and external occlusions from interaction. •Experiments on three public benchmarks demonstrate state-of-the-art performance against leading methods.
Wei_Fine-Grained_Classification_With_Noisy_Labels_CVPR_2023	Abstract Learning with noisy labels (LNL) aims to ensure model generalization given a label-corrupted training set. In this work, we investigate a rarely studied scenario of LNL on fine-grained datasets (LNL-FG), which is more practical and challenging as large inter-class ambiguities among fine-grained classes cause more noisy labels. We empiri- cally show that existing methods that work well for LNL fail to achieve satisfying performance for LNL-FG, aris- ing the practical need of effective solutions for LNL-FG. To this end, we propose a novel framework called stochas- tic noise-tolerated supervised contrastive learning (SNSCL) that confronts label noise by encouraging distinguishable representation. Specifically, we design a noise-tolerated supervised contrastive learning loss that incorporates a weight-aware mechanism for noisy label correction and se- lectively updating momentum queue lists. By this mecha- nism, we mitigate the effects of noisy anchors and avoid inserting noisy labels into the momentum-updated queue. Besides, to avoid manually-defined augmentation strategies in contrastive learning, we propose an efficient stochas- tic module that samples feature embeddings from a gener- ated distribution, which can also enhance the representa- tion ability of deep models. SNSCL is general and compati- ble with prevailing robust LNL strategies to improve their performance for LNL-FG. Extensive experiments demon- strate the effectiveness of SNSCL.	1. Introduction Learning from noisy labels [12, 13, 18, 21, 26, 40, 55, 58] poses great challenges for training deep models, whose per- formance heavily relies on large-scaled labeled datasets [28, 47–49]. Annotating data with high confidence would be resource-intensive, especially for some domains, such as medical and remote sensing images [29, 36, 37, 41, 46]. ∗denotes corresponding author Class 1Class 2Class 1Class 2 Class 2Class 1LNL - generic classificationRandom NoiseDependent NoiseLNL-FG - fine-grained classification Class 1Class 2 Figure 1. LNL-FG is more challenging than LNL on generic clas- sification. denote mislabeled samples. Thus, label noise would inevitably arise and then greatly degrade the generalization performance of deep models. Previous methods [1,6,7,9,18,23,38,53,54] in LNL al- ways focus on generic classification ( e.g.CIFAR-10 & 100) and artificially construct random label noise [21, 23, 42, 43] and dependent label noise [9, 18, 38, 53, 55] to evaluate the performance of their algorithms. In this work, we extend LNL to fine-grained classification, which is a rarely studied task. Firstly, this scenario is more realistic since annota- tors are easier to be misguided by indistinguishable charac- teristics among fine-grained images and give an uncertain target. Fig. 1 illustrates comparison between two types of noise simulated on generic and fine-grained sets. Fur- ther, we extensively investigate the performance of prevail- ing LNL methods on our proposed LNL-FG task. The de- tailed results are shown in Fig. 2. Although these robust al- gorithms lead to statistically significant improvements over vanilla softmax cross-entropy on LNL, these gains do not transfer to LNL-FG task. Instead, some methods degrade the generalization performance of deep models compared to cross-entropy. Intuitively, due to large inter-class ambi- guity among those classes in LNL-FG, the margin between noisy samples and the decision boundary in the fine-grained dataset is smaller than that in the generic dataset, leading to severe overfitting of deep models to noisy labels. De- spite this fact, the typical method for better representation, This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11651 5565 Co-teachingJoCoRMW-NetMLC SYM GCEConf. p Label sVanilla 62 57 72 JoCoRMW-NetMLC GCEConf. p Label s Vanilla60SYM6780 70 65 76 VanillaLabel s Conf. p GCEJoCoRSYM Co-teachingMLC MW-Net Co-teaching42DivideMix DivideMix 7380 Vanilla GCESYM Co-teachingJoCoRMW-NetMLCDivideMix Label s Conf. p 66DivideMixLNL-FG accuracy (%)Figure 2. Comparison results of previous methods on four fine-grained benchmarks with 20% random label noise. Methods with same color and shape belong to the same strategy. The X-axis denotes their performance on typical LNL tasks while the Y-axis denotes that on LNL-FG tasks. It is obvious that not all robust methods outperform the performance of vanilla cross-entropy on LNL-FG task. . More analysis and results can be found in Appx. A. i.e., DivideMix, consistently achieves better performance on both LNL and LNL-FG tasks (See Fig. 2). From this perspective, we consider that encouraging discrimitive fea- ture not only confronts overfitting to label noise but also facilitates the learning of fine-grained task. For this, contrastive learning (CL), as a powerful unsu- pervised learning approach for generating discrimitive fea- ture [4,8,11,14,31], has attracted our attention. CL methods usually design objective functions as supervised learning to perform pretext similarity measurement tasks derived from an unlabeled dataset, which can learn effective visual repre- sentations in downstream tasks, especially for fine-grained classification [3]. The following work, supervised con- trastive learning (SCL) [15], leverages label information to further enhance representation learning, which can avoid a vast training batch and reduce the memory cost. However, SCL cannot be directly applied to the noisy scenario as it is lack of noise-tolerated mechanism. To resolve the noise-sensitivity of SCL, we propose a novel framework named stochastic noise-tolerated su- pervised contrastive learning (SNSCL), which contains a noise-tolerated contrastive loss and a stochastic module. For the noise-tolerated contrastive loss, we roughly cate- gorize the noise-sensitive property of SCL into two parts of noisy anchors and noisy query keys in the momentum queue. To mitigate the negative effect introduced by noisy anchors or query keys, we design a weight mechanism for measuring the reliability score of each sample and give cor- responding weight. Based on these weights, we modify the label of noisy anchors in current training batch and selec- tively update the momentum queue for decreasing the prob- ability of noisy query keys. These operations are adaptive and can achieve a progressive learning process. Besides, to avoid manual adjustment of strong augmentation strategies for SCL, we propose a stochastic module for more com- plex feature transformation. In practice, this module gener- ates the probabilistic distribution of feature embedding. By sampling operation, SNSCL achieves better generalization performance for LNL-FG. Our contributions can be summarized as • We consider a hardly studied LNL task, dubbed LNL-FG and conduct empirical investigation to show that some existing methods in LNL cannot achieve satisfy- ing performance for LNL-FG. • We design a novel framework dubbed stochastic noise- tolerated supervised contrastive learning (SNSCL), which alters the noisy labels for anchor samples and selectively updates the momentum queue, avoiding the effects of noisy labels on SCL. • We design a stochastic module to avoid manually- defined augmentation, improving the performance of SNSCL on representation learning. • Our proposed SNSCL is generally applicable to pre- vailing LNL methods and significantly improves their performance on LNL-FG. Extensive experiments on four fine-grained datasets and two real-world datasets consistently demonstrate the state- of-the-art performance of SNSCL, and further analysis ver- ify its effectiveness.
Wang_Co-SLAM_Joint_Coordinate_and_Sparse_Parametric_Encodings_for_Neural_Real-Time_CVPR_2023	Abstract We present Co-SLAM, a neural RGB-D SLAM system based on a hybrid representation, that performs robust cam- era tracking and high-fidelity surface reconstruction in real time. Co-SLAM represents the scene as a multi-resolution hash-grid to exploit its high convergence speed and abil- ity to represent high-frequency local features. In addi- tion, Co-SLAM incorporates one-blob encoding, to encour- age surface coherence and completion in unobserved ar- eas. This joint parametric-coordinate encoding enables real-time and robust performance by bringing the best of both worlds: fast convergence and surface hole filling. Moreover, our ray sampling strategy allows Co-SLAM to perform global bundle adjustment over all keyframes in- stead of requiring keyframe selection to maintain a small number of active keyframes as competing neural SLAM ap- proaches do. Experimental results show that Co-SLAM runs at10−17Hz and achieves state-of-the-art scene reconstruc- tion results, and competitive tracking performance in vari- ous datasets and benchmarks (ScanNet, TUM, Replica, Syn- thetic RGBD). Project page: https://hengyiwang. github.io/projects/CoSLAM ⋆Indicates equal contribution.	1. Introduction Real-time joint camera tracking and dense surface re- construction from RGB-D sensors has been a core problem in computer vision and robotics for decades. Traditional SLAM solutions exist that can robustly track the position of the camera while fusing depth and/or color measurements into a single high-fidelity map. However, they rely on hand- crafted loss terms and do not exploit data-driven priors. Recent attention has turned to learning-based models that can exploit the ability of neural network architectures to learn smoothness and coherence priors directly from data. Coordinate-based networks have probably become the most popular representation, since they can be trained to predict the geometric and appearance properties of any point in the scene in a self-supervised way, directly from images. The most notable example, Neural Radiance Fields (NeRF) [14], encodes scene density and color in the weights of a neural network. In combination with volume rendering, NeRF is trained to re-synthesize the input images and has a remark- able ability to generalize to nearby unseen views. Coordinate-based networks embed input point coordi- nates into a high dimensional space, using sinusoidal or other frequency embeddings, allowing them to capture high-frequency details that are essential for high-fidelity geometry reconstruction [1]. Combined with the smooth- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13293 COORDINATE PARAMETRIC JOINT REFERENCE Frequency DenseGrid DenseGrid+OneBlob NICE-SLAM [42] OneBlob HashGrid HashGrid+OneBlob (Ours) GT Figure 2. Illustration of the effect of different encodings on completion. COORDINATE based encodings achieve hole filling but require long training times. PARAMETRIC encodings allow fast training, but fail to complete unobserved regions. JOINT coordinate and para- metric encoding (Ours) allows smooth scene completion and fast training. NICE-SLAM [42] uses a dense parametric encoding. ness and coherence priors inherently encoded in the MLP weights, they constitute a good choice for sequential track- ing and mapping [26]. However, the weakness of MLP- based approaches is the long training times required (some- times hours) to learn a single scene. For that reason, re- cent real-time capable SLAM systems built on coordinate networks with frequency embeddings such as iMAP [26] need to resort to strategies to sparsify ray sampling and re- duce tracking iterations to maintain interactive operation. This comes at the cost of loss of detail in the reconstruc- tions which are oversmoothed (Fig. 5) and potential errors in camera tracking. Optimizable feature grids, also known as parametric embeddings, have recently become a powerful alternative scene representation to monolithic MLPs, given their ability to represent high-fidelity local features and their extremely fast convergence (orders of magnitude faster) [7, 10, 15, 32, 40]. Recent efforts focus on sparse alternatives to these parametric embeddings such as octrees [28], tri-plane [2], hash-grid [15] or sparse voxel grid [12, 13] to improve the memory efficiency of dense grids. While these represen- tations can be fast to train and are therefore well suited to real-time operation, they fundamentally lack the smooth- ness, and coherence priors inherent to MLPs and strug- gle with hole-filling in areas without observation. NICE- SLAM [42] is a recent example of a multi-resolution fea- ture grid-based SLAM method. Although it does not suffer from over-smoothness and captures local detail (as shown in Fig. 2) it cannot perform hole-filling which might in turn lead to drift in camera pose estimation. Our first contribution is to design a joint coordinate and sparse grid encoding for input points that brings together the benefits of both worlds to the real-time SLAM frame-work. On the one hand, the smoothness and coherence pri- ors provided by coordinate encodings (we use one-blob [16] encoding), and on the other hand the optimization speed and local details of sparse feature encodings (we use hash grid [15]), resulting in more robust camera tracking and high-fidelity maps with better completion and hole filling. Our second contribution relates to the bundle adjustment (BA) step in the joint optimization of the map and cam- era poses. So far, all neural SLAM systems [26, 42] per- form BA using rays sampled from a very small subset of selected keyframes. Restricting the optimization to a very small number of viewpoints results in decreased robustness in camera tracking and increased computation due to the need for a keyframe-selection strategy. Instead, Co-SLAM performs global BA, sampling rays from all past keyframes, which results in an important boost in robustness and per- formance in pose estimation. In addition, we show that our BA optimization requires a fraction of the iterations of NICE-SLAM [42] to achieve similar errors. In practice, Co- SLAM achieves SOTA performance in camera tracking and 3D reconstruction while maintaining real time performance. Co-SLAM runs at 15-17Hz on Replica and Syn- thetic RGB-D datasets [1], and 12-13Hz on ScanNet [5] and TUM [25] scenes — faster than NICE-SLAM (0.1- 1Hz) [42] and iMAP [26]. We perform extensive evaluations on various datasets (Replica [24], Synthetic RGBD [1], ScanNet [6], TUM [25]) where we outperform NICE-SLAM [42] and iMAP [26] in reconstruction and achieve better or at least on-par tracking accuracy.
Wei_Super-Resolution_Neural_Operator_CVPR_2023	Abstract We propose Super-resolution Neural Operator (SRNO), a deep operator learning framework that can resolve high- resolution (HR) images at arbitrary scales from the low- resolution (LR) counterparts. Treating the LR-HR image pairs as continuous functions approximated with different grid sizes, SRNO learns the mapping between the corre- sponding function spaces. From the perspective of ap- proximation theory, SRNO ﬁrst embeds the LR input into a higher-dimensional latent representation space, trying to capture sufﬁcient basis functions, and then iteratively ap- proximates the implicit image function with a kernel inte- gral mechanism, followed by a ﬁnal dimensionality reduc- tion step to generate the RGB representation at the target coordinates. The key characteristics distinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efﬁciently implemented via the Galerkin- type attention, which possesses non-local properties in the spatial domain and therefore beneﬁts the grid-free contin- uum; and 2) the multilayer attention architecture allows for the dynamic latent basis update, which is crucial for SR problems to “hallucinate” high-frequency information from the LR image. Experiments show that SRNO outperforms existing continuous SR methods in terms of both accuracy and running time. Our code is at https://github.com/ 2y7c3/Super-Resolution-Neural-Operator .	1. Introduction Single image super-resolution (SR) addresses the in- verse problem of reconstructing high-resolution (HR) im- ages from their low-resolution (LR) counterparts. In a data- driven way, deep neural networks (DNNs) learn the inver- sion map from many LR-HR sample pairs and have demon- strated appealing performances [4, 20,21,24,29,40,41]. Nevertheless, most DNNs are developed in the conﬁgura- tion of single scaling factors, which cannot be used in sce- *Equal contributions. This work is supported by the National Natural Science Foundation of China (61871055). †Corresponding author. ℒ𝒦𝒫 Set5( ) DIV2K( ) Urban100( )ℋ(Ωℎ𝑐) ℋ(Ωℎ𝑓)⋅⋅⋅ ⋅⋅⋅ ⋅⋅⋅𝒱(Ωℎ𝑓) SRNOFigure 1. Overview of Super-Resolution Neural operator (SRNO). SRNO is composed of three parts, L(Lifting),K(ker- nel integrals) andP(Projection), which perform consecutively to learn mappings between approximation spaces H( hc)and H( hf)associated with grid sizes hcandhf, respectively. The key component,K, uses test functions in the latent Hilbert space V( hf)to seek instance-speciﬁc basis functions. narios requiring arbitrary SR factors [37, 38]. Recently, im- plicit neural functions (INF) [5, 18] have been proposed to represent images in arbitrary resolution, and paving a feasi- ble way for continuous SR. These networks, as opposed to storing discrete signals in grid-based formats, represent sig- nals with evaluations of continuous functions at speciﬁed coordinates, where the functions are generally parameter- ized by a multi-layer perceptron (MLP). To share knowl- edge across instances instead of ﬁtting individual functions for each signal, encoder-based methods [5, 15,26] are pro- posed to retrieve latent codes for each signal, and then a de- coding MLP is shared by all the instances to generate the required output, where both the coordinates and the cor- responding latent codes are taken as input. However, the point-wise behavior of MLP in the spatial dimensions re- sults in limited performance when decoding various objects, particularly for high-frequency components [30, 32]. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18247 Neural operator is a newly proposed neural network ar- chitecture in the ﬁeld of computational physics [17, 19,23] for numerically efﬁcient solvers of partial differential equa- tions (PDE). Stemming from the operator theory, nueral op- erators learn mappings between inﬁnite-dimensional func- tion spaces, which is inherently capable of continuous func- tion evaluations and has shown promising potentials in var- ious applications [9, 13,27]. Typically, neural operator con- sists of three components: 1) lifting, 2) iterative kernel in- tegral, and 3) projection. The kernel integrals operate in the spatial domain, and thus can explicitly capture the global relationship constraining the underlying solution function of the PDE. The attention mechanism in transformers [36] is a special case of kernel integral where linear transforms are ﬁrst exerted to the feature maps prior to the inner prod- uct operations [17]. Tremendous successes of transformers in various tasks [6, 22,34] have shown the importance of capturing global correlations, and this is also true for SR to improve performance. [8]. In this paper, we propose the super-resolution neural op- erator (SRNO), a deep operator learning framework that can resolve HR images from their LR counterparts at ar- bitrary scales. As shown in Fig.1, SRNO learns the map- ping between the corresponding function spaces by treating the LR-HR image pairs as continuous functions approxi- mated with different grid sizes. The key characteristics dis- tinguishing SRNO from prior continuous SR works are: 1) the kernel integral in each layer is efﬁciently implemented via the Galerkin-type attention, which possesses non-local properties in the spatial dimensions and have been proved to be comparable to a Petrov-Galerkin projection [3]; and 2) the multilayer attention architecture allows for the dy- namic latent basis update, which is crucial for SR problems to “hallucinate” high-frequency information from the LR image. When employing same encoders to capture features, our method outperforms previous continuous SR methods in terms of both reconstruction accuracy and running time. In summary, our main contributions are as follows: • We propose the methodology of super-resolution neu- ral operator that maps between ﬁnite-dimensional function spaces, allowing for continuous and zero-shot super-resolution irrespective the discretization used on the input and output spaces. • We develop an architecture for SRNO that ﬁrst ex- plores the common latent basis for the whole training set and subsequently reﬁnes an instance-speciﬁc basis by the Galerkin-type attention mechanism. • Numerically, we show that the proposed SRNO outper- forms existing continuous SR methods with less run- ning time, and even generates better results on the res- olutions for which the ﬁxed scale SR networks were trained.
Wu_PointConvFormer_Revenge_of_the_Point-Based_Convolution_CVPR_2023	Abstract We introduce PointConvFormer, a novel building block for point cloud based deep network architectures. Inspired by generalization theory, PointConvFormer combines ideas from point convolution, where filter weights are only based on relative position, and Transformers which utilize feature- based attention. In PointConvFormer, attention computed from feature difference between points in the neighborhood is used to modify the convolutional weights at each point. Hence, we preserved the invariances from point convolu- tion, whereas attention helps to select relevant points in the neighborhood for convolution. PointConvFormer is suitable for multiple tasks that require details at the point level, such as segmentation and scene flow estimation tasks. We exper- iment on both tasks with multiple datasets including Scan- Net, SemanticKitti, FlyingThings3D and KITTI. Our re- sults show that PointConvFormer offers a better accuracy- speed tradeoff than classic convolutions, regular transform- ers, and voxelized sparse convolution approaches. Visual- izations show that PointConvFormer performs similarly to convolution on flat areas, whereas the neighborhood selec- tion effect is stronger on object boundaries, showing that it has got the best of both worlds. The code will be available.	1. Introduction Depth sensors for indoor and outdoor 3D scanning have significantly improved in terms of both performance and affordability. Hence, their common data format, the 3D point cloud, has drawn significant attention from academia and industry. Understanding the 3D real world from point clouds can be applied to many application domains, e.g. robotics, autonomous driving, CAD, and AR/VR. However, unlike image pixels arranged in regular grids, 3D points are unstructured, which makes applying grid based Convolu- tional Neural Networks (CNNs) difficult. Various approaches have been proposed in response to this challenge. [3, 5, 31, 35, 37, 65] introduced interesting ways to project 3D point clouds back to 2D image space *this work was done entirely at Apple Inc., Wenxuan Wu was an intern at Apple Inc. when he participated in the work Input Grid SizeMethodRuntime (ms)# Params (M)mIoU(%)10cmMinkowskiNet52.937.960.7PointConv23.45.462.6FastPointTransformer140.037.966.5PointConvFormerPointConvFormer-Lite41.929.45.51.671.470.65cmMinkowskiNet73.537.967.0PointConv36.75.468.5FastPointTransformer147.737.970.0PointConvFormerPointConvFormer-Lite59.650.55.51.973.773.32cmMinkowskiNetPointConv115.688.837.99.372.270.3FastPointTransformer312.037.972.5Stratified Transformer1689.318.874.3PointConvFormer145.59.474.5Figure 1. Performance vs. running time on ScanNet. PointCon- vFormer achieves a state-of-the-art 74.5% mIoU while being effi- cient with faster speed and way less learnable parameters. Larger dot indicates more learnable parameters. All results are reported on a single TITAN RTX GPU and apply 2D convolution. Another line of research di- rectly voxelizes the 3D space and apply 3D discrete con- volution, but it induces massive computation and memory overhead [47,62]. Sparse 3D convolution operations [9,20] save a significant amount of computation by computing convolution only on occupied voxels. Some approaches directly operate on point clouds [41, 57,58,64,69,81]. [57,58] are pioneers which aggregate in- formation on point clouds using max-pooling layers. Others proposed to reorder the input points with a learned transfor- mation [42], a flexible point kernel [69], and a convolutional operation that directly work on point clouds [75, 81] which utilizes a multi-layer perceptron (MLP) to learn convolu- tion weights implicitly as a nonlinear transformation from the relative positions of the local neighbourhood. The approach to directly work on points is appealing to us because it allows direct manipulation of the point co- ordinates, thus being able to encode rotation/scale invari- ance/equivariance directly into the convolution weights [41, 42, 58, 94]. These invariances serve as priors to make the models more generalizable. Besides, point-based ap- proaches require less parameters than voxel-based ones, which need to keep e.g. 3×3×3convolution kernels on all input and output channels. Finally, point-based ap- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 21802 proaches can utilize k-nearest neighbors (kNN) to find the local neighborhood, thus can adapt to variable sampling densities in different 3D locations. However, so far the methods with the best accuracy- speed tradeoff have still been the sparse voxel-based ap- proaches or a fusion between sparse voxel and point-based models. Note that no matter the voxel-based or point-based representation, the information from the input is exactly the same, so it is unclear why fusion is needed. Besides, fu- sion adds significantly to model complexity and memory usage. This leads us to question the component that is in- deed different between these representations: the general- ization w.r.t. the irregular local neighbourhood. The shape of the kNN neighbourhood in point-based approaches varies in different parts of the point cloud. Irrelevant points from other objects, noise and the background might be included in the neighborhood, especially around object boundaries, which can be detrimental to the performance and the ro- bustness of point-based models. To improve the robustness of models with kNN neigh- borhoods, we refer back to the generalization theory of CNNs, which indicates that points with significant fea- ture correlation should be included in the same neighbor- hood [40]. A key idea in this paper is that feature corre- lation can be a way to filter out irrelevant neighbors in a kNN neighborhood, which makes the subsequent convolu- tion more generalizable. We introduce PointConvFormer, which computes attention weights based on feature differ- ences and uses that to reweight the points in the neighbor- hood in a point-based convolutional model, which indirectly “improves” the neighborhood for generalization. The idea of using feature-based attention is not new, but there are important differences between PointConvFormer and other vision transformers [16,54,96]. PointConvFormer combines features in the neighborhood with point-wise con- volution, whereas Transformer attention models usually adopt softmax attention in this step. In our formulation, the positional information is outside the attention, hence viewpoint-invariance can be introduced into the convoulu- tional weights. We believe that invariance helps generaliz- ing across neighborhood (size/rotation) differences between training/testing sets, especially with a kNN neighborhood. We evaluate PointConvFormer on two point cloud tasks, semantic segmentation and scene flow estimation. For semantic segmentation, experiment results on the indoor ScanNet [11] and the outdoor SemanticKitti [2] demon- strate superior performances over classic convolution and transformers with a much more compact network. The per- formance gap is the most significant at low resolutions, e.g. on ScanNet with a 10cm resolution we achieved more than 10% improvement over MinkowskiNet with only 15% of its parameters (Fig. 1). We also apply PointConvFormer as the backbone of PointPWC-Net [82] for scene flow esti-mation, and observe significant improvements on FlyingTh- ings3D [48] and KITTI scene flow 2015 [52] datasets as well. These results show that PointConvFormer could po- tentially compete with sparse convolution as the backbone choice for dense prediction tasks on 3D point clouds.
Wu_NewsNet_A_Novel_Dataset_for_Hierarchical_Temporal_Segmentation_CVPR_2023	Abstract Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granu- larity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the ex- isting fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos †Equal Contribution Corresponding Authors: Bing Li (bing.li@kaust.edu.sa) and Ruizhi Qiao (ruizhiqiao@tencent.com).in over 900 hours, associated with several tasks for hierar- chical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with exten- sive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is pub- licly available at https://github.com/NewsNet- Benchmark/NewsNet .	1. Introduction Temporal video segmentation is a critical problem in video understanding, which is essential for many video ap- plications such as video classification [10,15,37,58,59,62], This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10669 Table 1. Data comparison between the NewsNet and other related datasets. The NewsNet provides various multimodal data and hierarchical temporal segmentation annotations. Doc: documentary, Ads: advertisement. ( Please refer to our project page for more details. ) Dataset # VideoDurationModality# Annotation(s) per VideoSource(hours) Topic Story Scene Event A VS [75] 197 - Visual - - - 14.2 Ads BBC [6] 11 9 Visual - - 49.7 - Doc OVSD [48] 21 10 Visual - - 28.9 - Generic Kinetics-GEBD [51] 54,691 152 Audio + Visual - - - 4.9 Action MovieNet [23] † 1,100 2174 Text + Audio + Visual - - 66.0 849.1 Movie RAI [7] 10 - Visual - - - 98.7 News TI-News [35] 477 244 Audio + Visual - 55.6 - 530.4 News NewsNet (Ours) 1,000 946 Text + Audio + Visual 8.5 51.6 87.9 654.4 News † The number of annotations for Scene andShot is counted from MovieScene [47], which is a subset of MovieNet. captioning [3, 34, 63, 71] and retrieval [5, 16, 31, 52]. Tem- poral video segmentation aims to group successive video frames into short segments along the temporal dimension. With the explosive growth of long-form videos, it is desir- able that temporal video segmentation can convert a video into more meaningful segments for more efficient access to the video. However, it is challenging to develop effective temporal segmentation tools for long-form videos, since this requires a comprehensive understanding of video struc- ture, while long-form videos contain complex content. Towards temporal video segmentation, existing works explore shot, event, and scene segmentation tasks, respec- tively. Shot segmentation [38, 55, 57] divides a video into shots, where a shot consists of consecutive and visually continuous frames captured by a camera without interrup- tion [23]. Yet, shot segmentation only considers low-level visual cues ( i.e. visual similarity), lacking semantic under- standing. Instead, event segmentation [25, 51, 56] divides a video by detecting the moments of changes such as ac- tion/subject changes. To better capture the underlying struc- ture of a video, recent works [12, 47, 65] introduce video scene segmentation which segments a video into scene seg- ments, each comprising successive shots semantically re- lated to the same scene. Scene segmentation enables a coarser and higher-level representation than shot segmen- tation. However, compared to the rich content of mas- sive long-form videos, scene/event is fine-grained and often lacks a high-level summarization of video content, which is insufficient for capturing the complex semantic structure of many videos and briefly representing video content. In this work, we first explore how to comprehensively represent the complex structure of a long-form video for temporal video segmentation. Humans can hierarchically divide a video into segments of different granularities ac- cording to multi-level semantic information ( e.g.scene and topic), from the perspective of cognitive science. Natu- ral language processing researchers have widely exploredtopic-level understanding [29, 41] for summarizing docu- ments [8, 42], while little effort has been devoted to long- form videos. Inspired by these observations, besides scene and event, we propose to introduce two higher-level seman- tics ( i.e. story and topic) into temporal video segmentation, to provide a brief and semantic structure representation. As a result, such hierarchical and multi-level understand- ing brings about scalable video structure representation for temporal video segmentation on long-form videos. That is, a long-form video can be split into finer segments with lower-level semantics ( e.g. scene), but also can be summa- rized into coarser ones yet with higher-level semantics ( e.g. topic) by recursively grouping finer segments, which com- prehensively represents video structures from coarse to fine. However, the community lacks high-quality datasets to conduct this research. In particular, as shown in Table 1, most datasets only provide temporal structure annotations with regard to events or scenes. TI-News [35] and Movi- eScene [47] provide two levels of annotations, but these datasets lack topic-level ones. To effectively break this limitation, we build a novel large-scale dataset for hierarchical temporal segmentation, named NewsNet. The unique properties of our NewsNet introduce many advantages. First, it is among the largest datasets in the news domain. We collect over 900 hours of videos from 20 mainstream news platforms. It has a highly diverse distribution of data. Second, we carefully anno- tated it frame-by-frame with 4 hierarchical levels to ensure its quality can meet our needs. Third, it is multimodal, in- cluding textual, visual, and audio information. Due to the nature of the news, the alignment across modalities is ac- curate, which makes multimodal joint training of models feasible. Finally, the videos in NewsNet provide a complete understanding of public events. Compared with other video datasets [4,23,26], it introduces more objective open-world knowledge ( e.g., news introduction) while including sub- jective factual commentary ( e.g., host comments on news 10670 events), making it more amenable to real-life application. Based on NewsNet, we empirically highlight two promising directions for long-span temporal segmentation: 1) Infusing Multi-Modality knowledge can significantly im- prove the performance of long-form temporal segmentation; 2) Although story- and topic-level segmentation is challeng- ing, it can be benefited from hierarchical modeling with the event- and scene-level segmentation tasks. The main contributions of this paper are as follows: • We propose a novel large-scale dataset NewsNet for long-form video structure understanding. This dataset is derived from 900+ hours of video and annotated with 4 hierarchical levels of semantics. • NewsNet provides dense annotations and multi-modal information, promoting diverse benchmarks: sep- arate/hierarchical temporal video segmentation in scene/story/topic levels, as well as other common tasks like classification, video localization/grounding, and highlight detection. • We formulate a new benchmark, i.e.,hierarchical mod- eling in the temporal segmentation task, which needs a single model to predict segments of multiple hierarchi- cal levels. Based on the empirical study, we bring in- sights into how hierarchical modeling potentially ben- efits the temporal video segmentation task, which was almost never discussed.
Williams_Black-Box_Sparse_Adversarial_Attack_via_Multi-Objective_Optimisation_CVPR_2023	Abstract Deep neural networks (DNNs) are susceptible to adver- sarial images, raising concerns about their reliability in safety-critical tasks. Sparse adversarial attacks, which limit the number of modified pixels, have shown to be highly effective in causing DNNs to misclassify. However, exist- ing methods often struggle to simultaneously minimize the number of modified pixels and the size of the modifica- tions, often requiring a large number of queries and as- suming unrestricted access to the targeted DNN. In con- trast, other methods that limit the number of modified pixels often permit unbounded modifications, making them easily detectable.To address these limitations, we propose a novel multi-objective sparse attack algorithm that efficiently min- imizes the number of modified pixels and their size during the attack process. Our algorithm draws inspiration from evolutionary computation and incorporates a mechanism for prioritizing objectives that aligns with an attacker’s goals. Our approach outperforms existing sparse attacks on CIFAR-10 and ImageNet trained DNN classifiers while requiring only a small query budget, attaining competitive attack success rates while perturbing fewer pixels. Over- all, our proposed attack algorithm provides a solution to the limitations of current sparse attack methods by jointly minimizing the number of modified pixels and their size. Our results demonstrate the effectiveness of our approach in restricted scenarios, highlighting its potential to enhance DNN security.	1. Introduction Although deep neural networks (DNNs) have made im- pressive strides in computer vision tasks [17, 21, 22, 24, 28, 37, 38, 48], recent work has shown that small optimized perturbations to input images can cause DNNs to misclas- sify [1, 18, 26, 31, 34, 42]. As adversarial images have been found to exist in the physical world [26, 27, 43], particu- Proposed Method SSIM = 0 :9977Sparse-RS SSIM = 0 :8667 SSIM = 0 :9906 SSIM = 0 :9632Figure 1. This illustration shows adversarial images and their cor- responding perturbations generated by two different algorithms, the proposed method and Sparse-RS [10], both attacking an adver- sarially trained CIFAR-10 [19] (top) and ImageNet [35] (bottom). While both images are adversarial, the perturbation generated by the Sparse-RS algorithm visibly distorts the image, whereas the proposed method’s adversarial image remains more similar to the original. This similarity is demonstrated by calculating the struc- tural similarity (SSIM) between the adversarial images and the original. The effectiveness of the Sparse-RS algorithm is there- fore questionable due to the significant distortion it causes. lar concern has been expressed on their impact on security- critical applications [1]. To address this issue, previous works have emphasized the importance of generating strong adversarial images [3]. As a result, significant effort has been devoted to developing effective attack methods that This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12291 can construct perturbations capable of causing DNN classi- fiers to misclassify images while preserving their semantic content. Most adversarial attack methods in the literature formu- late the attack as an optimization problem where a loss function is minimized to achieve the desired misclassifica- tion of an image. While many attack methods in the lit- erature constrain the adversarial perturbation by its l2or l∞norm [2, 4, 6–8, 18, 23, 26, 29, 36, 42, 45, 46] and allow all pixels of an image to be perturbed, there is also a need to develop sparse attack methods that constrain adversarial perturbations by their l0norm [10,11,15,16,30,41,44,47]. Such adversarial images have been found to also exist in the physical world and have shown to be as effective as the more traditional l2orl∞-constrained adversarial images. Numerous sparse attack methods have been proposed to address both the white-box [11, 15, 16, 44, 47] and black- box scenarios [10, 30, 41]. In the white-box scenario, the attacker has full access to a DNN’s information, while the black-box scenario assumes the attack only has access to the outputted class probabilities. In this work, we focus on the black-box scenario. While existing attack meth- ods have shown success in generating adversarial images, they often struggle to handle the trade-off between optimiz- ing the loss function and minimizing the perturbations lp norms, where p= 0,1,2or∞. Several methods only con- strain the number of perturbed pixels [10, 11, 41], allow- ing the size of the perturbation to be unbounded. Despite their efficiency, the unbounded nature of the generated per- turbations result in obvious distortions of the original im- age, as shown in Fig. 1. On the other hand, other meth- ods allow the l0norm to be minimized along with the loss function [15,16,47], while constraining the perturbation by itsl2orl∞norm. These methods either generate an ad- versarial perturbation and then reduce its l0norm [47] or add an l0norm penalty term to the optimized loss func- tion [15, 16]. Despite their good performance, these meth- ods only address the white-box scenario and assume access to a large number of DNN queries, limiting their applicabil- ity to query-limited scenarios [23]. Therefore, by not prop- erly handling this trade-off, existing methods are limited in their applicability and effectiveness in real-world scenarios. Within the evolutionary computation field a classic ap- proach to handling conflicting objectives is the use of a domination relation [13] which characterises the trade-off between objectives and is used to compare solutions within a population-based evolutionary algorithm. While the orig- inal approach assigns equal weight to each objective, the domination mechanism can be adapted to reflect the at- tacker’s preferences. In this work, our goal is to generate sparse adversarial perturbations with low l0andl2norms in an efficient manner. Our contributions can be summarised as follows:• To address the challenge of generating sparse adver- sarial perturbations, we formulate the problem as a bi- objective optimization problem. By constraining the perturbation to a set of discrete values, we show that minimizing of the l2norm also minimizes the l0norm. • We propose a new dominance relation to compare so- lutions that gives first priority to minimizing the loss function and then to minimizing its l2norm. • To generate adversarial perturbations, we propose a population-based heuristic attack method that utilizes two distributions to generate new solutions. These dis- tributions mimic the crossover and mutation operators commonly used in evolutionary computation. By sam- pling from these distributions, our approach explores the search space in an efficient and effective manner. • To evaluate the effectiveness of our approach, we con- duct attacks on DNN classifiers trained on the CIFAR- 10 and ImageNet datasets, using a low-query budget and considering both targeted and non-targeted attack scenarios. Our empirical results demonstrate that our proposed method outperforms state-of-the-art white- box sparse attacks, as well as the black-box Sparse-RS attack method [10], in terms of success rate and num- ber of perturbed pixels.
Xiao_Masked_Images_Are_Counterfactual_Samples_for_Robust_Fine-Tuning_CVPR_2023	Abstract Deep learning models are challenged by the distribu- tion shift between the training data and test data. Re- cently, the large models pre-trained on diverse data have demonstrated unprecedented robustness to various distri- bution shifts. However, fine-tuning these models can lead to a trade-off between in-distribution (ID) performance and out-of-distribution (OOD) robustness. Existing methods for tackling this trade-off do not explicitly address the OOD ro- bustness problem. In this paper, based on causal analysis of the aforementioned problems, we propose a novel fine- tuning method, which uses masked images as counterfactual samples that help improve the robustness of the fine-tuning model. Specifically, we mask either the semantics-related or semantics-unrelated patches of the images based on class activation map to break the spurious correlation, and refill the masked patches with patches from other images. The resulting counterfactual samples are used in feature-based distillation with the pre-trained model. Extensive experi- ments verify that regularizing the fine-tuning with the pro- posed masked images can achieve a better trade-off between ID and OOD performance, surpassing previous methods on the OOD performance. Our code is available at https: //github.com/Coxy7/robust-finetuning .	1. Introduction Deep learning has made impressive advances in various tasks on computer vision. Despite the remarkable perfor- mance achieved on benchmark datasets, deep models are often challenged by the distribution shift between the train- ing data and test data [5,15,16,33]. It is commonly assumed that the training and test samples follow the same distribu- tion, which may not hold in real-world applications due to the unpredictable change of lighting conditions, viewpoints, backgrounds, etc. Although there are attempts to improve the robustness of deep models to the distribution shift (or OOD robustness), it is still rather under-explored [29, 38]. *Corresponding author. CAM -based Masking Pre-trained Robust ModelVanilla Fine -tuned ModelRobust Fine -tuned Model (ours)What is the semantics? Red admiral butterfly. A bunch of flowers. Flowers. learning fromFigure 1. Illustration of our work. Vanilla fine-tuned models tend to learn spurious correlations that degrade the OOD robustness. To tackle this issue, our model learns from the pre-trained model on the counterfactual CAM-based masked images. Recently, large-scale models pre-trained on diverse data have demonstrated unprecedented robustness to various dis- tribution shifts [8, 19, 32]. Fine-tuning these pre-trained models on downstream tasks can be a promising approach to building robust models for different applications. How- ever, it is found that while fine-tuning improves the per- formance on in-distribution (ID) data, it may reduce that on out-of-distribution (OOD) data [23, 43]. To tackle this trade-off, several methods [23, 42, 43] have been proposed to improve both ID and OOD performance in fine-tuning. However, they do not explicitly address the OOD robust- ness problem; instead, they implicitly preserve the robust- ness of the pre-trained model by constraining the distortion of pre-trained weights or using model ensembles. In this paper, we revisit the issue of robustness degrada- tion in fine-tuning from a causal perspective. A large-scale pre-trained model somewhat shows properties in causality and stays robust to OOD samples [41]. However, when fine-tuning on downstream tasks, a majority of the model parameters tend to be adjusted for the downstream task in fine-tuning due to the highly entangled representation of This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 20301 images, arguably destructive to the generalizable knowl- edge [20, 35]. In contrast, distribution shifts are usually sparse in the underlying causal factorization of the data gen- eration process [7, 35]. In this low-dimensional case, if we know which variables vary with different data distributions in this factorization ( i.e., the non-stationary factors), we can achieve the OOD robustness by simply excluding their in- fluence on the final predictions of the model. Specifically, we consider a Structural Causal Model (SCM) [31] for the object-centric image generation process, as depicted in Fig. 2. In this SCM, images are generated according to a non-stationary domain-relevant factor and a stationary semantic factor. Between them is a spurious correlation caused by a hidden non-stationary confounder that influences how the domain-relevant factor changes with the semantic one. To retain the OOD robustness, a fine- tuning model should avoid mapping non-stationary domain- relevant features to the predicted semantics. To this end, we propose to fine-tune the models with masked images, which serve as counterfactual samples breaking the spurious correlation. Training on these sam- ples helps preserve the stationary and generalizable knowl- edge of the pre-trained model. Concretely, we either mask the patches that contribute most to the label ( i.e., the main object) or mask those with the least contribution ( e.g., the context), which can be implemented based on class activa- tion map (CAM) [9,46]. Such image masking forms an ma- nipulation of a factual image and produces a counterfactual sample. Since the pre-trained model can better disentan- gle invariant features across domains, we require the fine- tuning model to learn from the pre-trained model on these counterfactual samples, as illustrated in Fig. 1. Further- more, we argue that simply dropping the masked patches may be insufficient to alleviate the risk of fitting spurious correlations, and we propose to refill the masked patches with those from other images. We study different combinations of masking strategies (e.g., masking the object or the context) and refilling strate- gies ( e.g., filling with patches from single or multiple im- ages). Experimental results suggest that most of the strate- gies are applicable to the construction of counterfactual samples that help improve the robustness in fine-tuning, while masking the object generally achieves the best ro- bustness. Compared with existing methods [23, 42, 43] on fine-tuning CLIP [32] models, our approach achieves bet- ter average accuracy on various OOD datasets without re- lying on model ensembles or weight constraints. We also find that, taking the weight-space ensemble of the zero-shot model and our fine-tuned model following WiSE-FT [43], hardly improves the trade-off between ID and OOD accu- racy, which contradicts previous observations and implies that our approach may produce essentially different models in comparison with conventional fine-tuning.
Wang_Non-Line-of-Sight_Imaging_With_Signal_Superresolution_Network_CVPR_2023	Abstract Non-line-of-sight (NLOS) imaging aims at reconstruct- ing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various ﬁelds, it has drawn much attention in recent years. However, long ex- posure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imag- ing quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolu- tion signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful recon- structions of the hidden scene under both confocal and non- confocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing opti- cal systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging.	1. Introduction Non-line-of-sight (NLOS) imaging problem usually em- ploys a high temporal resolution optical system to recover the hidden object around the corner. As shown in Fig. 1a, photons emitted by the laser are collected by the detector after three diffuse reﬂections: the reﬂection at the visible wall, the reﬂection at the hidden object, and the reﬂection at the visible wall again. Scanning a region on the visible wall can get measured histogram data to recover the hidden scene. Due to its potential applications in various ﬁelds, such as auto-driving, disaster relief, and remote sensing, the NLOS imaging problem has become an emerging ﬁeld since it was ﬁrst proposed by Kirmani et al. [17] in 2009. (a) NLOS scenario (b) Scanning points (c) Results Figure 1. A typical NLOS scenario and reconstruction results. (a) An illustration of a typical NLOS scenario. (b) An illustration of the scanning points on the visible wall. Using the proposed pipeline, only the blue circle points are needed to be illuminated, and the signal at the yellow square points can be recovered with the proposed pipeline. (c) Comparisons of the reconstruction re- sults of the pyramid with different signals. The maximum intensity projection of the albedo values along the depth direction is shown in the upper left corner (GT). The results are reconstructed by the original signal with spatial resolution 3232(Original), the sub- sampled signal with spatial resolution 88(Low), and the signal recovered by the proposed network with spatial resolution 3232 (Ours). Many methods have been proposed to improve the prac- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17420 Figure 2. Flowchart of the proposed pipeline. The high resolution signal is recovered from the low resolution signal using a neural network. The hidden scene is then reconstructed using state-of-the-art imaging algorithms designed for high resolution signal. ticability [1, 3, 13, 14, 34, 35, 39] or reconstruction quality [10, 11, 26, 41]. A back-projection method with Laplacian of Gaussian ﬁlter (LOG-BP) [18] is introduced to improve the reconstruction quality of the back-projection method [38]. In 2018, O’Toole et al. [31] ﬁrst apply the light-cone- transform (LCT) method to image the hidden object with a confocal scanning mode. Young et al. [45] then propose the directional light-cone-transform (D-LCT) method, which simultaneously recovers the albedo and surface normal of the hidden object. From the perspective of wave charac- teristics, Lindell et al . [22] introduce the fast frequency- wavenumber migration (F-K) method to NLOS. Another re- markable imaging method proposed by Liu et al. [25] em- ploys the phasor ﬁeld model [9, 33] and provides an efﬁ- cient solution for fast NLOS imaging under non-confocal scenario [24]. Further
Xie_MAESTER_Masked_Autoencoder_Guided_Segmentation_at_Pixel_Resolution_for_Accurate_CVPR_2023	Abstract Accurate segmentation of cellular images remains an elusive task due to the intrinsic variability in morphology of biological structures. Complete manual segmentation is unfeasible for large datasets, and while supervised meth- ods have been proposed to automate segmentation, they often rely on manually generated ground truths which are especially challenging and time consuming to generate in biology due to the requirement of domain expertise. Fur- thermore, these methods have limited generalization capac- ity, requiring additional manual labels to be generated for each dataset and use case. We introduce MAESTER (Masked AutoEncoder guided SegmenTation at pixEl Resolution), a self-supervised method for accurate, subcellular structure segmentation at pixel resolution. MAESTER treats segmen- tation as a representation learning and clustering problem. Specifically, MAESTER learns semantically meaningful to- ken representations of multi-pixel image patches while si- multaneously maintaining a sufficiently large field of view for contextual learning. We also develop a cover-and-stride inference strategy to achieve pixel-level subcellular struc- ture segmentation. We evaluated MAESTER on a publicly available volumetric electron microscopy (VEM) dataset of primary mouse pancreatic islets βcells and achieved up- wards of 29.1%improvement over state-of-the-art under the same evaluation criteria. Furthermore, our results are competitive against supervised methods trained on the same tasks, closing the gap between self-supervised and super- vised approaches. MAESTER shows promise for alleviating the critical bottleneck of ground truth generation for imaging related data analysis and thereby greatly increasing the rate of biological discovery. Code available at https : / / github . com / bowang-lab/MAESTER *Equal contribution †Project lead ‡Co-senior author	1. Introduction Imaging is widely used in biology to study the organi- zation, morphology, and function of cells and subcellular structures [13, 26, 28, 31, 32]. Segmentation of structures and objects of interest in the acquired images is often crit- ical for downstream analysis. Recent innovations in high throughput imaging technology enables larger scale datasets to be collected more quickly and cost efficiently [23, 24, 32]. Scalable and accurate segmentation hence becomes a crucial bottleneck to overcome. For example, volumetric electron microscopy (VEM) can generate terabytes of imaging data in a single run, enabling biologists to uncover ultrastructural features of cells at unprecedented resolution and scale in 3D [24]. Manual segmentation of such datasets are unfea- sible and especially when substantial domain knowledge is required for annotation of structures captured in the imaging volume. With recent advancements in the field of machine learn- ing, automatic methods involving convolutional neural net- works (CNNs) have been developed to aid the segmentation process to great success [8, 18]. However, these methods often require extensive manual labels to train in the first place. Furthermore, supervised models often exhibit lim- ited generalization capacity, necessitating additional ground truth generation efforts for each new dataset or use case. Presently, there is a dire need for a self-supervised segmenta- tion method to bypass the initial bottleneck of manual label generation, particularly when the cost and time of acquir- ing training supervision far exceeds the capacity to generate unlabelled data. In addition to being self-supervised, the method needs to incorporate a few inductive biases to tackle the challenges of biological image segmentation. First, the texture of objects from the same class often remains consistent, despite great variability in shapes and sizes that cellular structures can exhibit. Therefore, the model needs to learn semantically meaningful representation of small image patches belonging to each structure of interest and distinguish between different textures. Second, the model needs to be capable of producing This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3292 Figure 1. MAESTER achieves self-supervised representation learning and segmentation through: (a) patchifying large sample from EM imaging (b) learning patch-level representation through predicting randomly masked region, (c) inferring the representation for the center voxel of each patch, (d) showing the 3D-rendered volume of our MAESTER generated segmentation, (e) demonstrating cover-and-stride strategy. features that precisely correspond to small regions in the original image. Not only will this increase the resolution of the resulting segmentation, it will also allow the method to take advantage of the locality assumption, which posits that small groups of adjacent pixels are more likely to belong to the same class. Third, the model needs to be context aware. While distinguishing between multi-pixel patches of images alone can achieve good segmentation results [5], wehypothesize that including a greater field of view (FOV) as context is crucial for better representation learning for the purpose of subcellular structure segmentation. Transformer based architectures have seen recent suc- cesses in computer vision [3, 15, 29]. The token-wise rep- resentation of image patches offers a natural way to inject inductive bias into our self-supervised segmentation model. We introduce MAESTER (Masked AutoEncoder guided Seg- 3293 menTation at pixEl Resolution), a self-supervised method that can achieve accurate, pixel-level segmentation of sub- cellular structures. MAESTER works in two stages. During training, MAESTER takes as input a large FOV ( F×F pixels) containing ample local context which is further bro- ken down into multi-pixel patches of size P×Ppixels. The choice of Pis sufficiently small to allow each patch to be treated as a single class under the locality assump- tion while achieving higher spatial precision. The attention mechanism of a vision transformer (ViT) [3] encoder then allows information sharing between nearby patches. Fur- thermore, taking inspiration from the Masked Autoencoder (MAE) [6] learning paradigm, we incorporate the surrogate task of multi-pixel patch masking and reconstruction via a light weight ViT decoder for each sampled FOV of a given image to simultaneously learn semantically meaningful to- ken representations of all patches in the FOV . During infer- ence, we deploy the trained encoder to generate millions of representations of unlabelled image patches via a novel cover-and-stride inference strategy. These representations are then clustered to produce a desired number of classes for self-supervised segmentation, leading to the final segmenta- tion of the given VEM dataset. To our knowledge, we are the first to use the transformer architecture to incorporate the inductive biases needed for self-supervised subcellular structure segmentation. We also repurposed and optimized the MAE learning paradigm for generating semantically relevant token representations of multi-pixel sized image patches for classification into bio- logically concordant clusters for segmentation rather than for pretraining or representation learning at the image level. Lastly, we introduce a cover-and-stride inference strategy to achieve pixel level segmentation of the given biological images. We tested MAESTER on the betaSeg dataset [20], consisting of primary mouse pancreatic islets βcells and yielded upwards of 29.1% increase in performance compared to prior state-of-the-art [5]. We also benchmarked against Segmenter [27] and vanilla ViT [3], two supervised seg- mentation models with access to all ground truth labels in addition to the raw images used to train MAESTER. We find MAESTER achieved competitive results for the predomi- nant classes, closing the gap between supervised and self- supervised segmentation models. We believe MAESTER has the potential to drastically speed up the experimental cycle of biological imaging experiments by alleviating the critical bottleneck of manual label generation and greatly increasing the rate of scientific inquiry in cell biology.
Xu_Constructing_Deep_Spiking_Neural_Networks_From_Artificial_Neural_Networks_With_CVPR_2023	Abstract Spiking neural networks (SNNs) are well-known as brain-inspired models with high computing efficiency, due to a key component that they utilize spikes as information units, close to the biological neural systems. Although spik- ing based models are energy efficient by taking advantage of discrete spike signals, their performance is limited by cur- rent network structures and their training methods. As dis- crete signals, typical SNNs cannot apply the gradient de- scent rules directly into parameter adjustment as artificial neural networks (ANNs). Aiming at this limitation, here we propose a novel method of constructing deep SNN mod- els with knowledge distillation (KD) that uses ANN as the teacher model and SNN as the student model. Through the ANN-SNN joint training algorithm, the student SNN model can learn rich feature information from the teacher ANN model through the KD method, yet it avoids training SNN from scratch when communicating with non-differentiable spikes. Our method can not only build a more efficient deep spiking structure feasibly and reasonably but use few time steps to train the whole model compared to direct train- ing or ANN to SNN methods. More importantly, it has a superb ability of noise immunity for various types of ar- tificial noises and natural signals. The proposed novel method provides efficient ways to improve the performance of SNN through constructing deeper structures in a high- throughput fashion, with potential usage for light and effi- cient brain-inspired computing of practical scenarios.	1. Introduction By referring to the information processing mechanism and structural characteristics of the biological nervous system, spiking neural networks (SNNs) are remarkably *Corresponding authors: jrshen@zju.edu.cn and gpan@zju.edu.cn.good at computational intelligence tasks [26] and suitable for processing unstructured information, with stronger au- tonomous learning capabilities and ultra-low power con- sumption [2, 7, 23, 37]. Although various engineering effort has been made in this area, such type of biological information processing system still underperforms artificial systems (artificial neu- ral networks, ANNs) in some common computer tasks, such as image classification. One possible reason for this is that typical SNNs lack deep hierarchical network struc- tures compared to those from ANNs. Due to the non- differentiable spikes, typical SNNs are restricted to global training rules which lead to various of current SNNs being just shallow fully-connected layer based [28, 34]. Limited by training rules and structures, although SNNs can signifi- cantly handle spatial-temporal data efficiently, it is difficult to train a deep SNN directly as using backpropagation (BP) in ANNs do [22]. By drawing on some key tricks from ANNs, some stud- ies want to improve image classification accuracy which is led by SNNs by combing structure and learning rules those has been proven to be effective in improving model performance in ANNs. [3, 6] proposed methods to con- vert ANNs to SNNs by controlling the output and network structure of ANN and SNN to be as consistent as possi- ble. Through this way, although they can build effective deep SNNs, these conversion methods suffer long training time and lack some intermediate information during ANN training period. [12,19] tried to adopt the threshold of spik- ing neurons to make them suitable for using gradient sur- rogate method, these models adopted too complex neuron models to get good performance and take up large computa- tional memory and cost. [8] did interesting work on directly converting an adjusted ANN to an SNN using the theoret- ical equivalence between activation and firing rate, which achieves superior performance. [29] constructed ANN-SNN hybrid models to improve the feature extraction, but these This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7886 hybrid models suffered a difficult training process. Aiming at constructing efficient SNNs, this paper pro- posed a brand-new method using knowledge distillation (KD) to let student models (SNNs) absorb rich informa- tion from teacher models (ANNs). KD [4] can transfer the knowledge of one network to another network, two net- works can be homogeneous or heterogeneous. This is done by training a teacher network and then using the output of the teacher network and the real tag of the data to train the student network. KD can be used to transform a network from a large network to a small network, retaining perfor- mance close to that of a large network, or to transfer knowl- edge learned from multiple networks into a single network, making the performance of a single network close to the results of Ensemble. Under the guidance of teacher models, the wanted SNNs model can be trained in a layer-wise manner [20]. Unlike traditional ANN-SNN conversion requires the same model structure of two models, the proposed KD conversion can make a heterogeneous network structure of them, for ex- ample, if the teacher ANN is larger and deeper, the student SNN can be smaller and shallower. This kind of KD conver- sion provides sufficient flexibility to construct any arbitrary SNN. In this paper, we propose a novel KD based training method to construct deep SNNs which avoids restricting corresponding network structures between ANN and SNN during the training period. Through a unified ANN-SNN loss function, we can construct the SNN model from well- trained ANN, accelerate the training time and save mem- ory usage. We adopt supervised gradient surrogate method as basic student SNN training rules. We evaluated the proposed method on several image classification datasets (MNIST, CIFAR10, and CIFAR100) and their noisy vari- ations. Experimental results showed that the proposed method can get pretty good image classification perfor- mance with a light SNN model. The main contributions are as follows: • This paper proposed a KD based conversion method to construct deep SNNs from ANNs, which only takes less training latency and allows the structure of the SNN and ANN to be heterogeneous. • Through the proposed ANN-SNN joint training method, the student SNN model can absorb more in- formation from ANN during training method, com- pared to offline ANN-SNN conversion, the proposed method significantly helped to improve the perfor- mance of the student SNN model. • We demonstrate the efficiency and effectiveness of the proposed distilling SNN method through evaluations of several datasets and noisy ones. Experimental re- sults show that we can construct a more reasonableSNN which can achieve state-of-the-art performance on experimental datasets with less training latency and show better anti-noise ability.
Wang_Pixels_Regions_and_Objects_Multiple_Enhancement_for_Salient_Object_Detection_CVPR_2023	Abstract Salient object detection (SOD) aims to mimic the human visual system (HVS) and cognition mechanisms to identify and segment salient objects. However, due to the complex- ity of these mechanisms, current methods are not perfect. Accuracy and robustness need to be further improved, par- ticularly in complex scenes with multiple objects and back- ground clutter. To address this issue, we propose a novel approach called Multiple Enhancement Network (MENet) that adopts the boundary sensibility, content integrity, it- erative refinement, and frequency decomposition mecha- nisms of HVS. A multi-level hybrid loss is firstly designed to guide the network to learn pixel-level, region-level, and object-level features. A flexible multiscale feature enhance- ment module (ME-Module) is then designed to gradually aggregate and refine global or detailed features by chang- ing the size order of the input feature sequence. An iter- ative training strategy is used to enhance boundary fea- tures and adaptive features in the dual-branch decoder of MENet. Comprehensive evaluations on six challenging benchmark datasets show that MENet achieves state-of-the- art results. Both the codes and results are publicly available at https://github.com/yiwangtz/MENet. *The corresponding authors	1. Introduction Salient object detection (SOD) aims to identify the most visually conspicuous regions in an image that are consis- tent with the human visual system (HVS) and cognition mechanisms [9,13,40]. SOD can eliminate redundant infor- mation and improve computational performance for many high-level computer vision tasks, such as action recognition [4, 60], image segmentation [3, 33], image captioning [52], object tracking [14], and video summary [58]. Fully con- volutional networks (FCNs) [25] based SOD models have been particularly effective at improving SOD performance in recent years [40]. However, accurately segmenting com- plex object boundaries remains a challenging task for SOD. This is especially true when the geometry and/or boundaries of these objects are complex, or when scenes are chaotic or cluttered [9, 10], as shown in Fig. 1. An intuitive solution for addressing this problem is to explore the mechanisms of the human vision system (HVS) [55] and some of which have been used to improve SOD models, as described below. (i) A human tends to enhance recognition by alternating between viewing the entire object and the details of complex scenes, which has been utilized for various visual tasks [16, 17, 22, 36]. (ii) HVS is sen- sitive to both boundary/contour and structural information, so dual-branch feature refinement structures have been de- veloped to incorporate extra-edge information to enhance salient feature learning [11, 22, 24, 43, 47, 53, 57]. Some This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10031 Figure 1. Illustration of MAE (left part) and some visual results (right part) for the proposed MENet with some recent state-of- the-art SOD methods: EDN [45], AADFNet [57], SAC [16], and ICON [59]. Please refer to Sec.5 for detailed experimental set- tings. The MENet model achieves the lowest MAE score with the most precise and complete boundaries. structural similarity measurements (e.g., Structural Simi- larity Index (SSIM) [41]) and regional similarity measure- ments (e.g., Intersection over Union (IoU) [35] and Dice [12]) are also adopted by SOD models [32, 42, 43, 47, 59] in the loss functions. (iii) Human vision is indeed holistic and continuous so that it perceives objects and scenes as or- ganized wholes [18], which are composed of parts that are meaningful and coherent in relation to each other. ICON [59] proposes to improve the integrity from both macro- and micro-level perspectives by enhancing integrity infor- mation hidden in channels of features. EDN [45] employs a powerful down-sampling technique to learn a global view of the whole image effectively. (iv) According to the hu- man visual spatial frequency model [30], an image can be decomposed into or synthesized by high-spatial frequency and low-spatial frequency parts. As a starting point in this work, we intend to use the mechanisms outlined above to further improve SOD performance for complex scenes. In this work, we propose a multi-enhancement network (MENet) that effectively integrates the above HVS mecha- nisms in a U-Net-like [37] encoder-decoder framework to produce more accurate SOD for complex scenes. Foremost, MENet employs the image frequency decomposition idea to design a two-stream feature learning decoder for boundaries (high frequencies) and inner body regions (low frequen- cies). This setting is different from the existing two-branch (or edge-aware) methods [11, 22, 47, 51, 53, 56, 57] that use one branch for the boundary and the other one for the entire object, such as EGNet [53] and AFNet [11]. Particularly, there is no interaction between the intermediate features of the two branches of MENet, so it reduces the interference of inaccurate boundary information with global features. This is because boundary features need to be highly discrimi- native against the background, while global features need consistency and robustness. Although LDF [43] also learns internal regional features in one branch, its detailed map and body map cannot be computed accurately and efficiently forgeometrically complex objects. Then, we propose an iterative training strategy to pro- gressively enhance features by alternately aggregating high- and low-level features to mimic HVS bottom-up and top- down refinement mechanisms. To produce high- and low- level features flexibly, we design a multiscale feature en- hancement module (ME-Module) as the core of each branch by leveraging atrous spatial pyramid pooling (ASPP) [34] and global-local attention [6]. In addition, we introduce the HVS holistic and continu- ous mechanism to loss function design. We present a multi- level hybrid loss, which evaluates the pixel-, region-, and object-level similarities between predicted saliency maps and ground-truth (GT) saliency maps. For pixel-level loss, we also use Binary Cross Entropy (BCE) [7] loss to ensure network accuracy and convergence speed. As for region- level loss, we divide a saliency map into four sub-regions of equal size and then calculate the sum of weighted re- gional similarities through SSIM and IoU. Then, inspired by SSIM and S-measure [5], an object-level loss is designed by the contrast and distribution statistics of the foreground between the GT map and the predicted map. A similar hy- brid loss is reported in BASNet [32], but it uses a simple combination of BCE, IoU, and SSIM for the whole saliency map without partitioning regions. Following is a summary of our contributions. • We propose to leverage not only pixel-level but also region-level and object-level similarity measures in loss to increase prediction accuracy and integrity, and then design a multi-level hybrid loss to implement this proposal. • We design a multiscale feature enhancement module (ME-Module) to mimic HVS bottom-up and top-down refinement mechanisms. ME-Module can gradually propagate and produce comprehensive global or de- tailed features by changing the size and order of the input features. • We propose a novel Multiple Enhancement Network (MENet) for dealing with SOD in complex scenes by integrating multiple HVS mechanisms into the net- work structure and loss function. Specifically, a two- branch decoder equipped with ME-Modules is de- signed to incrementally refine the boundary and adap- tive features by an iterative training strategy and the proposed multilevel hybrid loss. The results of quantitative and qualitative experiments on six datasets demonstrate MENet outperforms the state- of-the-art methods by a large margin, as shown in Fig. 1. 10032 Figure 2. Illustration of the overall architecture and the pipeline of the MENet.
Xu_H2ONet_Hand-Occlusion-and-Orientation-Aware_Network_for_Real-Time_3D_Hand_Mesh_Reconstruction_CVPR_2023	Abstract Real-time 3D hand mesh reconstruction is challenging, especially when the hand is holding some object. Beyond the previous methods, we design H2ONet to fully exploit non-occluded information from multiple frames to boost the reconstruction quality. First, we decouple hand mesh recon- struction into two branches, one to exploit finger-level non- occluded information and the other to exploit global hand orientation, with lightweight structures to promote real- time inference. Second, we propose finger-level occlusion- aware feature fusion, leveraging predicted finger-level oc- clusion information as guidance to fuse finger-level infor- mation across time frames. Further, we design hand-level occlusion-aware feature fusion to fetch non-occluded infor- mation from nearby time frames. We conduct experiments on the Dex-YCB and HO3D-v2 datasets with challenging hand-object occlusion cases, manifesting that H2ONet is able to run in real-time and achieves state-of-the-art per- formance on both the hand mesh and pose precision. The code will be released on GitHub.	1. Introduction Estimating 3D hand meshes from RGB images is a fun- damental task useful for many applications, e.g., augmented reality [17, 52], behavior understanding [26, 44], etc. To support these applications, user experience is very impor- tant, so the reconstruction should be accurate and robust, as well as fast, i.e., real-time. Despite the promising results achieved by the recent works, it is still very challenging to simultaneously meet all the requirements, particularly when the hand is severely occluded, e.g., holding some object. Several recent methods are proposed for 3D hand mesh reconstruction from a single RGB image [13, 15, 31, 38– 42, 45]. To alleviate the negative effect of occlusion, some try to extract occlusion-robust features by adopting the spa- Figure 1. Structural comparison between our H2ONet and previ- ous methods. We decouple 3D hand mesh reconstruction into two branches, one to reconstruct the hand mesh at canonical pose M and the other to regress the global hand orientation R, such that we can fuse finger- and hand-level occlusion-aware features from multiple frames to better exploit the non-occluded information. tial attention mechanism applied in 3D hand pose estima- tion [14, 65, 68]. When the amount of occlusion is small, focusing more on non-occluded regions can help improve the network performance. However, the performance would largely drop when the occluded regions dominate, implying that relying solely on the prior information of hand shape and pose is insufficient. Besides, the attention mechanism brings extra computation and memory overhead. Though recent methods [8, 52] adopt lightweight frameworks for real-time inference, the influence of occlusion is ignored. On the other hand, some recent works [18,57] start to ex- plore multi-frame RGB images as input for 3D hand mesh reconstruction. SeqHAND [57] integrates LSTM as a fea- ture extractor to memorize the hand motion over consecu- tive frames. Liu et al. [35] constrain the smoothness of hand shape and pose by designing inter-frame losses. Yet, they This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17048 do not have specific designs to explicitly deal with the oc- clusion. Hasson et al. [18] leverages an optical-flow-guided strategy to promote photometric consistency. Nevertheless, the extra information is limited, as they use only two adja- cent frames. Also, though multi-frame inputs provide more information, it is non-trivial to effectively extract and fuse multi-frame features to improve the reconstruction quality. In this paper, we present H2ONet , aHand-Occlusion- and-Orientation-aware Network, aiming at exploiting non- occluding information from multi-frame images to recon- struct the 3D hand mesh. Our goal is to meet the require- ments of (i) effectively utilizing the inter-frame information and (ii) explicitly alleviating the interference of occlusion. First, as the hand orientation information and hand shape information are mixed in feature space, it is hard to directly fuse features from multiple frames. To better exploit use- ful information, we decouple hand mesh reconstruction into two tasks: one for hand mesh reconstruction at the canon- ical pose and the other for hand orientation regression, as shown in Fig. 1. The key advantages are that we can better fuse multi-frame features without considering hand orien- tation differences, and it enables us to apply strategies to alleviate the ill-posed issue in estimating hand orientation. Second, to handle self and object occlusions on the hand, we propose to exploit non-occluding information spatially across fingers and temporally across frames. For the for- mer, we design finger-level occlusion-aware feature fusion that leverages predicted finger-level occlusion probabilities to guide the adaptive fusion of per-finger features from mul- tiple frames. For the latter, we design hand-level occlusion- aware feature fusion that catches auxiliary global informa- tion over frames guided by the hand-level occlusions. In summary, our main contributions are: • We design the hand-occlusion-and-orientation-aware network named H2ONet with a two-branch architec- ture to efficiently and effectively exploit non-occluding information from multiple frames. • We formulate finger-level occlusion-aware feature fu- sion and hand-level occlusion-aware feature fusion modules. The former aggregates non-occluded finger- level information from multiple frames to promote hand shape reconstruction, whereas the latter alleviates the ill-posed issue when estimating the global hand ori- entation in case the hand is temporarily occluded. • Through qualitative and quantitative comparisons on two datasets with severe hand occlusions, we show that H2ONet achieves state-of-the-art performance.
Wong_Heat_Diffusion_Based_Multi-Scale_and_Geometric_Structure-Aware_Transformer_for_Mesh_CVPR_2023	Abstract Triangle mesh segmentation is an important task in 3D shape analysis, especially in applications such as digi- tal humans and AR/VR. Transformer model is inherently permutation-invariant to input, which makes it a suitable candidate model for 3D mesh processing. However, two main challenges involved in adapting Transformer from nat- ural languages to 3D mesh are yet to be solved, such as i) extracting the multi-scale information of mesh data in an adaptive manner; ii) capturing geometric structures of mesh data as the discriminative characteristics of the shape. Current point based Transformer models fail to tackle such challenges and thus provide inferior performance for dis- cretized surface segmentation. In this work, heat diffusion based method is exploited to tackle these problems. A novel Transformer model called MeshFormer is proposed, which i) integrates Heat Diffusion method into Multi-head Self- Attention operation (HDMSA) to adaptively capture the fea- tures from local neighborhood to global contexts; ii) ap- plies a novel Heat Kernel Signature based Structure Encod- ing (HKSSE) to embed the intrinsic geometric structures of mesh instances into Transformer for structure-aware processing. Extensive experiments on triangle mesh seg- mentation validate the effectiveness of the proposed Mesh- Former model and show signiﬁcant improvements over cur- rent state-of-the-art methods.	1. Introduction Discretized surface semantic segmentation is a task to semantically classify the labeling of each discrete element in 3D discretized surface. Such discrete element can be tri- angle face in mesh input [22, 24, 37] or 3D point in point cloud input [11, 20, 26, 28, 34, 48–50, 52]. It is an essen- tial task in many applications for 3D vision and computer graphics, such as 3D human body analysis, digital humans and AR/VR, etc. In this work, we mainly focus on mesh representation as input, as point cloud representation canbe regarded as the special case of mesh which discards the surface connectivity. The challenges in learning mesh representation involves its inherent characteristics such as irregularity and un- orderedness. Following the success in NLP [7, 15, 44] and 2D computer vision domain [16, 29, 43, 46], Transformer model such as [20, 32, 50, 52] has been adopted as an effec- tive model for processing 3D point cloud input due to its in- herent capability in processing unordered point sets. Along such direction, it is natural to adapt the Transformer model to the mesh input, which is also one of the most common representation for 3D input modality. However, adapting Transformer model from natural languages to mesh input, with respect to the speciﬁc characteristics of mesh struc- tures, involves many challenges, such as i) extracting the multi-scale information of mesh data in an adaptive man- ner; ii) capturing geometric structures of mesh representa- tion as the discriminative characteristics. Sufﬁciently cap- turing such two essential information is the prerequisite for accurate mesh based semantic segmentation task. In ret- rospect to the recent point cloud based Transformer mod- els [20, 32, 52], it is found that all these methods did not provide effective approaches to tackle the challenges men- tioned above, and thus provided a limited increment in seg- mentation accuracy for mesh input. Recent work Swin Transformer [29] applies multiple ﬁxed-size windows for limiting the attention computation along scale-varying regions, and hence provides a hierar- chical Transformer model to extract multi-scale informa- tion for dense prediction task. However, such approach which uses ﬁxed-size windows is only applicable to regu- lar 2D images input, while it is infeasible for irregular input such as 3D meshes with diverse shapes. In order to adapt the Transformer model for adaptively extracting multi-scale information for irregular mesh input, it is essential to ex- tend its core operation, self-attention, to have capability in progressively comparing the feature similarity from local neighborhood to global range of the mesh input, in the form of intrinsic geometry of surface [9]. In fact, several seminal works [12,13] had studied on using heat diffusion method to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 4413 intrinsically communicate the interactions with the neigh- bouring vertices on discretized surface. Such heat diffusion method is able to compute the geodesic distance on mesh in a stable and accurate manner, which is essential to capture the intrinsic locality in discretized surface. In this work, we opt for the heat diffusion method to propose a novel self-attention operation called Heat Dif- fusion based Multi-head Self-Attention (HDMSA), which adaptively limits the self-attention computation within mul- tiple heat diffusion ranges to capture the multi-scale surface features from local neighborhood to global contexts. This extension facilitates the construction of multi-scale Trans- former encoder in the proposed MeshFormer model. The second challenging issue in adapting Transformer model for mesh input is to encode the geometric structural information of mesh as a supplement for shape-speciﬁc in- ductive bias of Transformer model. In Non-Euclidean do- main, very recent works such as SAN [25], GNN-LSPE [17] methods have investigated on injecting the eigenfunctions, which are derived from graph Laplacian operator, into posi- tional encoding as a kind of graph-speciﬁc inductive bias to traditional Transformer model. This approach exploits the spectral information of graph Laplacian to capture the struc- ture of input modality, and thus is considered as a promis- ing candidate for processing mesh input. However, this ap- proach suffers from the issue of eigenfunction sign ambigu- ity, that means eigenfunction with either positive or negative sign still satisﬁes the original eigenproblem and is associ- ated with the same eigenvalue, which lowers the discrimi- native power in the extracted structural information. In this work, instead of directly applying the spectral information, a novel Heat Kernel Signature based Structure Encoding (HKSSE) module is proposed, which effectively captures the intrinsic geometric structural information of the mesh while bypassing the issue of eigenfunction sign ambiguity. Moreover, it provides a more powerful way to capture the more advanced geometric information, i.e., the symmetry in geometry structure, which is very common in human body shapes. As a result, this capability brought to Transformer model reinforces a structure-aware segmentation prediction for mesh input. To the best of our knowledge, the proposed model is the ﬁrst mesh based Transformer model which integrates heat diffusion methods to tackle the discretized surface semantic segmentation problem. The main contributions of this work are summarized as follows: 1. With the heat diffusion extension, the proposed multi- head self-attention operation allows intrinsic commu- nication for vertices on mesh input from local neigh- borhood to global context and thus is able to capture multi-scale mesh features. 2. A novel heat kernel signature based structure encodingmodule is applied to embed the mesh intrinsic geomet- ric structures into Transformer for providing structure- aware segmentation output. 3. The proposed work demonstrates the feasibility on the extension of generic Transformer model structure for 3D mesh input with heat diffusion methods.
Wang_Task_Difficulty_Aware_Parameter_Allocation__Regularization_for_Lifelong_Learning_CVPR_2023	Abstract Parameter regularization or allocation methods are ef- fective in overcoming catastrophic forgetting in lifelong learning. However, they solve all tasks in a sequence uni- formly and ignore the differences in the learning difficulty of different tasks. So parameter regularization methods face significant forgetting when learning a new task very dif- ferent from learned tasks, and parameter allocation meth- ods face unnecessary parameter overhead when learning simple tasks. In this paper, we propose the Parameter Allocation & Regularization (PAR) , which adaptively select an appropriate strategy for each task from parameter allo- cation and regularization based on its learning difficulty. A task is easy for a model that has learned tasks related to it and vice versa. We propose a divergence estimation method based on the Nearest-Prototype distance to mea- sure the task relatedness using only features of the new task. Moreover, we propose a time-efficient relatedness-aware sampling-based architecture search strategy to reduce the parameter overhead for allocation. Experimental results on multiple benchmarks demonstrate that, compared with SOTAs, our method is scalable and significantly reduces the model’s redundancy while improving the model’s per- formance. Further qualitative analysis indicates that PAR obtains reasonable task-relatedness.	1. Introduction Recently, the lifelong learning [9] ability of neural net- works, i.e., learning continuously from a continuous se- quence of tasks, has been extensively studied. It is natural for human beings to constantly learn and accumulate knowl- edge from tasks and then use it to facilitate future learning. However, classical models [13, 18, 41] suffer catastrophic forgetting [12], i.e., the model’s performance on learned tasks deteriorates rapidly after learning a new one. To overcome the catastrophic forgetting, many param- *Corresponding author: Yin Zhang. Large carnivores Large omnivores and herbivoresReptiles Fruit and vegetables Parameter Regularization for the easy new task Current taskLearn the first task Parameter Allocation for the difficult new task ?Adaptive strategy ？ Adaptive Parameter Allocation & Regularization Based on T ask Learning Difficulty ModelLearning StrategyLabel Task Data 5-classes Classification T asksStrategies Model （Related to task ) （No related tasks before) （Related to ?) ExpertFigure 1. PAR adaptively selects a strategy from regularization and allocation to handle each task according to its learning diffi- culty. The difficulty depends not only on the task itself, but also on whether the model has previously learned tasks related to it. eter regularization or allocation methods have been pro- posed. Parameter regularization methods [10, 16, 19, 21, 23, 25, 27] alleviate forgetting by adding a regularization term to the loss function and perform well when the new task does not differ much from learned tasks. Parameter al- location methods based on static models [7, 15, 31, 36] and dynamic models [2,20,24,26,29,30,34,38,39,42,45,47] al- locate different parameters to different tasks and can adapt to new tasks quite different from learned tasks. However, the above methods solve all tasks in a sequence uniformly, and ignore the differences of learning difficulty of different tasks. This leads to the significant forgetting in parameter regularization methods when learning a new task which is quite different from learned tasks, and also leads to unnec- essary parameter cost in parameter allocation methods when learning some simple tasks. In this paper, we propose a difficulty-aware method Parameter Allocation & Regularization (PAR). As shown This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 7776 in Fig. 1, we assume that the learning difficulty of a task in continual learning depends not only on the task itself, but also on the accumulated knowledge in the model. A new task is easy to adapt for a model if it has learned re- lated tasks before and vice versa. Based on the assumption, the PAR adaptively adopts parameter allocation for difficult tasks and parameter regularization for easy tasks. Specif- ically, the PAR divides tasks into task groups and assigns each group a dedicated expert model. Given a new task, the PAR measures the relatedness between it and existing task groups at first. If the new task is related to one of the ex- isting groups, it is easy for the corresponding expert. The PAR adds the task to the related group and learns it by the expert via the parameter regularization. Otherwise, the new task is difficult for all existing experts, and the PAR assigns it to a new task group and allocates a new expert to learn it. There are two challenges in this work: the measurement of relatedness and the parameter explosion associated with parameter allocation. For the first one, we try to measure the relatedness by the KL divergence between feature distribu- tions of tasks. However, the KL divergence is intractable and needs to be estimated since the feature distributions of tasks are usually unknown. In addition, the constraint of lifelong learning that only data of the current task are avail- able exacerbates the difficulty of estimation. To solve above problems, inspired by the divergence estimation based on k-NN distance [44], we propose the divergence estimation method based on prototype distance, which only depends on the data of the current task. For the second one, we try to reduce parameter overhead per expert by searching com- pact architecture for it. However, the low time and memory efficiency is an obstacle to applying architecture search for a sequence of tasks in lifelong learning. To improve the efficiency of architecture search, we propose a relatedness- aware sampling-based hierarchical search. The main con- tributions of this work are as follows: • We propose a lifelong learning framework named Pa- rameter Allocation & Regularization (PAR), which se- lects an appropriate strategy from parameter allocation and regularization for each task based on the learning difficulty. The difficulty depends on whether the model has learned related tasks before. • We propose a divergence estimation method based on prototype distance to measure the distance between the new task and previous learned tasks with only data of the new task. Meanwhile, we propose a relatedness- aware sampling-based architecture search to reduce the parameter overhead of parameter allocation. • Experimental results on CTrL, Mixed CIFAR100 and F-CelebA, CIFAR10-5, CIFAR100-10, CIFAR100-20 and MiniImageNet-20 demonstrate that PAR is scal- able and significantly reduces the model redundancywhile improving the model performance. Exhaustive ablation studies show the effectiveness of components in PAR and the visualizations show the reasonability of task distance in PAR.
Wang_DR2_Diffusion-Based_Robust_Degradation_Remover_for_Blind_Face_Restoration_CVPR_2023	Abstract Blind face restoration usually synthesizes degraded low- quality data with a pre-defined degradation model for train- ing, while more complex cases could happen in the real world. This gap between the assumed and actual degra- dation hurts the restoration performance where artifacts are often observed in the output. However, it is expensive and infeasible to include every type of degradation to cover real-world cases in the training data. To tackle this robust- ness issue, we propose Diffusion-based Robust Degradation Remover (DR2) to first transform the degraded image to a coarse but degradation-invariant prediction, then employ an enhancement module to restore the coarse prediction to a high-quality image. By leveraging a well-performing de- noising diffusion probabilistic model, our DR2 diffuses in- put images to a noisy status where various types of degrada- tion give way to Gaussian noise, and then captures semantic information through iterative denoising steps. As a result, DR2 is robust against common degradation ( e.g. blur, re- size, noise and compression) and compatible with different †Corresponding author. *Cooperative Medianet Innovation Center.designs of enhancement modules. Experiments in various settings show that our framework outperforms state-of-the- art methods on heavily degraded synthetic and real-world datasets.	1. Introduction Blind face restoration aims to restore high-quality face images from their low-quality counterparts suffering from unknown degradation, such as low-resolution [5, 11, 27], blur [45], noise [23, 36], compression [10], etc. Great im- provement in restoration quality has been witnessed over the past few years with the exploitation of various facial pri- ors. Geometric priors such as facial landmarks [5], parsing maps [4, 5], and heatmaps [43] are pivotal to recovering the shapes of facial components. Reference priors [9,25,26] of high-quality images are used as guidance to improve details. Recent research investigates generative priors [39, 42] and high-quality dictionaries [14,24,48], which help to generate photo-realistic details and textures. Despite the great progress in visual quality, these meth- ods lack a robust mechanism to handle degraded inputs be- sides relying on pre-defined degradation to synthesize the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1704 training data. When applying them to images of severe or unseen degradation, undesired results with obvious artifacts can be observed. As shown in Fig. 1, artifacts typically ap- pear when 1) the input image lacks high-frequency informa- tion due to downsampling or blur ( 1strow), in which case restoration networks can not generate adequate information, or 2) the input image bears corrupted high-frequency in- formation due to noise or other degradation ( 2ndrow), and restoration networks mistakenly use the corrupted informa- tion for restoration. The primary cause of this inadaptabil- ity is the inconsistency between the synthetic degradation of training data and the actual degradation in the real world. Expanding the synthetic degradation model for training would improve the models’ adaptability but it is apparently difficult and expensive to simulate every possible degrada- tion in the real world. To alleviate the dependency on syn- thetic degradation, we leverage a well-performing denois- ing diffusion probabilistic model (DDPM) [16, 37] to re- move the degradation from inputs. DDPM generates im- ages through a stochastic iterative denoising process and Gaussian noisy images can provide guidance to the gen- erative process [6, 29]. As shown in Fig. 2, noisy images aredegradation-irrelevant conditions for DDPM genera- tive process. Adding extra Gaussian noise (right) makes dif- ferent degradation less distinguishable compared with the original distribution (left), while DDPM can still capture the semantic information within this noise status and re- cover clean face images. This property of pretrained DDPM makes it a robust degradation removal module though only high-quality face images are used for training the DDPM. Our overall blind face restoration framework DR2E con- sists of the Diffusion-based Robust Degradation Remover (DR2) and an Enhancement module. In the first stage, DR2 first transforms the degraded images into coarse, smooth, and visually clean intermediate results, which fall into a degradation-invariant distribution ( 4thcolumn in Fig. 1). In the second stage, the degradation-invariant images are fur- ther processed by the enhancement module for high-quality details. By this design, the enhancement module is com- patible with various designs of restoration methods in seek- ing the best restoration quality, ensuring our DR2E achieves both strong robustness and high quality. We summarize the contributions as follows. (1) We pro- pose DR2 that leverages a pretrained diffusion model to remove degradation, achieving robustness against complex degradation without using synthetic degradation for train- ing. (2) Together with an enhancement module, we em- ploy DR2 in a two-stage blind face restoration framework. The enhancement module has great flexibility in incorporat- ing a variety of restoration methods to achieve high restora- tion quality. (3) Comprehensive experiments show that our framework outperforms state-of-the-art methods on heavily degraded synthetic and real-world datasets. ▼ Laplace ✖ Gaussian 20 + JPEG ▲ Gaussian 10 ● Blur Original ▼ Laplace ✖ Gaussian 20 + JPEG ▲ Gaussian 10 ● Blur Add Noise σ σ Figure 2. Mean and standard variation of pixel-wise error dis- tribution. (Left) the error between original degraded input yand its ground truth low-resolution image ˆy(only bicubically down- sampled); (Right) the error between q(y500\|y)andq(ˆy500\|ˆy) sampled by Eq. (2), with extra Gaussian noise added by the dif- fusion function.
Wang_Cut_and_Learn_for_Unsupervised_Object_Detection_and_Instance_Segmentation_CVPR_2023	Abstract We propose Cut-and- LEaRn (CutLER), a simple ap- proach for training unsupervised object detection and seg- mentation models. We leverage the property of self- supervised models to ‘discover’ objects without supervision and amplify it to train a state-of-the-art localization model without any human labels. CutLER first uses our proposed MaskCut approach to generate coarse masks for multiple objects in an image, and then learns a detector on these masks using our robust loss function. We further improve performance by self-training the model on its predictions. Compared to prior work, CutLER is simpler, compatible with different detection architectures, and detects multiple objects. CutLER is also a zero-shot unsupervised detec- tor and improves detection performance AP 50by over 2.7 × on 11 benchmarks across domains like video frames, paint- ings, sketches, etc. With finetuning, CutLER serves as a low- shot detector surpassing MoCo-v2 by 7.3% APboxand 6.6% APmaskon COCO when training with 5% labels.	1. Introduction Object localization is a critical task in computer vision that enables AI systems to perceive, reason, plan and act inan object-centric manner. Training models for localization require special annotations like object boxes, masks, local- ized points, etc. which are both difficult and resource inten- sive to collect. Without accounting for overhead, annotating ∼164K images in the COCO dataset [32] with masks for just 80 classes took more than 28K human hours of annota- tion time. In this work, we study unsupervised object detec- tion and instance segmentation models that can be trained without any human labels. Our key insight is that simple probing and training mechanisms can amplify the innate lo- calization ability of self-supervised models [7], leading to state-of-the-art unsupervised zero-shot detectors. Our method Cut-and- LEaRn (CutLER) consists of three simple, architecture- and data-agnostic mechanisms. Con- sistent with prior self-supervised learning methods [7–9, 26], CutLER is trained exclusively on unlabeled ImageNet data without needing additional training data, but contrary to these methods, CutLER can be directly employed to per- form complex segmentation and detection tasks over a wide range of domains. First , we propose MaskCut that can au- tomatically produce multiple initial coarse masks for each image, using the pretrained self-supervised features. Sec- ond, we propose a simple loss dropping strategy to train detectors using the coarse masks while being robust to ob- jects missed by MaskCut. Finally , we observe that despite learning from these coarse masks, the detectors ‘clean’ the This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3124 ground truth and produce masks (and boxes) that are bet- ter than the coarse masks used to train them. Therefore, we further show that multiple rounds of self-training on the models’ own predictions allow it to evolve from capturing the similarity of local pixels to capturing the global geome- try of the object, thus producing finer segmentation masks. Prior work shows that a self-supervised vision trans- former (ViT) [15] can automatically learn patch-wise fea- tures that detect a single salient object in an image [7,38,43, 44,50]. However, unlike CutLER, such salient object detec- tion methods only locate a single, usually the most promi- nent, object and cannot be used for real world images con- taining multiple objects. While some recent methods, e.g., FreeSOLO [47] and DETReg [3], also aim at unsupervised multi-object detection (or multi-object discovery), they rely on a particular detection architecture, e.g., SOLO-v2 [48] or DDETR [5,54]. Additionally, apart from self-supervised features trained on ImageNet [12], the current state-of-the- art methods FreeSOLO and MaskDistill [42] also require ‘in-domain’ unlabeled data for model training. In contrast, CutLER works with various detection archi- tectures and can be trained solely on ImageNet, without requiring in-domain unlabeled data. Thus, during model training, CutLER does not see any images from any target dataset and yields a zero-shot model capable of detecting and segmenting multiple objects in diverse domains. Features of CutLER. 1) Simplicity: CutLER is simple to train and agnostic to the choice of detection and backbone architectures. Thus, it can be integrated effortlessly into existing object detection and instance segmentation works. 2) Zero-shot detector: CutLER trained solely on ImageNet shows strong zero-shot performance on 11 different bench- marks where it outperforms prior work trained with addi- tional in-domain data. We double the APbox 50performance on 10 of these benchmarks, as shown in Fig. 1, and even outperform supervised detectors on the UVO video instance segmentation benchmark. 3) Robustness: CutLER exhibits strong robustness against domain shifts when tested on im- ages from different domains such as video frames, sketches, paintings, clip arts, etc.4) Pretraining for supervised de- tection: CutLER can also serve as a pretrained model for training fully supervised object detection and instance seg- mentation models and improves performance on COCO, in- cluding on few-shot object detection benchmarks.
Wang_Two-Stream_Networks_for_Weakly-Supervised_Temporal_Action_Localization_With_Semantic-Aware_Mechanisms_CVPR_2023	Abstract Weakly-supervised temporal action localization aims to detect action boundaries in untrimmed videos with only video-level annotations. Most existing schemes detect tem- poral regions that are most responsive to video-level classi- fication, but they overlook the semantic consistency between frames. In this paper, we hypothesize that snippets with similar representations should be considered as the same action class despite the absence of supervision signals on each snippet. To this end, we devise a learnable dictionary where entries are the class centroids of the corresponding action categories. The representations of snippets identified as the same action category are induced to be close to the same class centroid, which guides the network to perceive the semantics of frames and avoid unreasonable localiza- tion. Besides, we propose a two-stream framework that in- tegrates the attention mechanism and the multiple-instance learning strategy to extract fine-grained clues and salient features respectively. Their complementarity enables the model to refine temporal boundaries. Finally, the developed model is validated on the publicly available THUMOS-14 and ActivityNet-1.3 datasets, where substantial experiments and analyses demonstrate that our model achieves remark- able advances over existing methods.	1. Introduction Temporal action localization (TAL) is committed to de- tecting action intervals in untrimmed videos. It has re- ceived increasing popularity recently due to its wide ap- plication in surveillance analysis, video summarization and retrieval [39, 44, 47], etc. Typically, fully-supervised TAL [51, 58, 59] is prohibitively expensive and unrealistic due to frame-level annotations, thus the weakly-supervised TAL (WS-TAL) [32, 34, 40, 45, 55] that only video-level annota- tions are required has been advocated recently. Most WS-TAL methods [8, 9, 31, 43, 45, 46] transform Figure 1. An example contains the action of “ Bicycle Motocross ” and the background. Representations depicted in monochromatic color are similar, which should be regarded to describe the same action (or background) and grouped together. The color brightness indicates the degree of similarity. localization into classification tasks that detect temporal regions contributing the most to video-level classification. They divide raw videos into fixed-length non-overlapping snippets, on which snippet-wise attention activations or class activation sequence (CAS) is generated. Temporal re- gions are detected by thresholding and merging these ac- tivations along the time dimension. Specifically, multiple instance learning (MIL) [36, 42] and the attention mecha- nism [11, 14, 53] are typically employed. The former ag- gregates snippets that are considered action instances with top-k confidence. However, such a regime over-emphasizes these snippets with top-k confidence, resulting in discard- ing potential clues in remaining snippets with lower confi- dence. Besides, MIL chooses the most discriminative snip- pets ignoring the completeness of action instances, which is incompatible with the localization task. Different from MIL, the attention-based mechanism independently yields class-agnostic confidence for each snippet, which is utilized as a weight to perform temporal pooling over all snippets and generate video-level representations for classification. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18878 Despite the usage of all snippet features for fine-grained patterns, class-agnostic confidence is semantically ambigu- ous and harmful to precise boundary detection. As a con- sequence, we design a two-stream network that integrates MIL and attention-based mechanisms to overcome their re- spective drawbacks. Then a late-fusion operation on the outputs of the two branches is conducted to acquire the final classification results. Furthermore, a crux of these works lies in accurately pre- dicting confidence scores that each snippet belongs to the foreground or background, which has a nontrivial impact on the subsequent boundary regression. Since the weakly- supervision paradigm does not provide explicit supervision signals, this problem becomes more intractable. A com- mon solution employs temporal class activation map [40] (TCAM) to discover snippets that respond to the video-level classification and assign them high confidence. Other alter- natives [11, 14, 45, 53] attempt to mitigate this problem by carefully formulating some attention generation and aggre- gation mechanisms. Nevertheless, these strategies neglect the semantic consistency between snippets. Intuitively, snippets with similar representations should be considered to be the same class despite the infeasibility of accessing snippet-level annotations. An example is also illustrated in Figure 1. We argue that it is unreasonable that there are no constraints to guarantee such a semantic relation. To address this intractable issue, we set a learnable dictionary where entries are class centroids of the corresponding action categories. The representations of snippets identified as the same action are induced to be close to the same class cen- troid. In this manner, the semantic relationship of snippets is explicitly explored to encourage a reasonable localization in the weakly-supervised paradigm. In a nutshell, the main contributions and innovations of this paper are summarized as follows: (1) A novel two- stream network that absorbs the merits of MIL and atten- tion mechanism is proposed to resolve WS-TAL. (2) To per- ceive semantic information, a learnable dictionary with eu- clidean constraint is designed to facilitate similar represen- tations to be considered as the same action class. (3) Ex- tensive experiments on THUMOS-14 and ActivityNet-1.3 benchmarks demonstrate that our model acquires remark- able advances. Besides, substantial ablation studies also re- veal that the proposed two-stream structure and semantic- aware modules are of effectiveness.
Wang_Deep_Learning_of_Partial_Graph_Matching_via_Differentiable_Top-K_CVPR_2023	Abstract Graph matching (GM) aims at discovering node match- ing between graphs, by maximizing the node- and edge- wise afﬁnities between the matched elements. As an NP- hard problem, its challenge is further pronounced in the existence of outlier nodes in both graphs which is ubiqui- tous in practice, especially for vision problems. However, popular afﬁnity-maximization-based paradigms often lack a principled scheme to suppress the false matching and resort to handcrafted thresholding to dismiss the outliers. This limitation is also inherited by the neural GM solvers though they have shown superior performance in the ideal no-outlier setting. In this paper, we propose to formulate the partial GM problem as the top- kselection task with a given/estimated number of inliers k. Speciﬁcally, we devise a differentiable top- kmodule that enables effective gradi- ent descent over the optimal-transport layer, which can be readily plugged into SOTA deep GM pipelines including the quadratic matching network NGMv2 as well as the linear matching network GCAN. Meanwhile, the attention-fused aggregation layers are developed to estimate kto enable automatic outlier-robust matching in the wild. Last but not least, we remake and release a new benchmark called IMC-PT-SparseGM, originating from the IMC-PT stereo- matching dataset. The new benchmark involves more scale- varying graphs and partial matching instances from the real world. Experiments show that our methods outperform other partial matching schemes on popular benchmarks.	1. Introduction The importance of graph matching (GM) is recognized by its successful applications in machine learning [ 26], molecule matching [ 47], and various vision tasks [ 15,30, ⇤Runzhong Wang and Ziao Guo contributed equally. Junchi Yan is the correspondence author who is also with Shanghai AI Laboratory. The work was in part supported by National Key Research and Development Program of China (2020AAA0107600), NSFC (62222607, U19B2035), Shanghai Committee Science and Technology Project (22511105100). IMC-PT-SparseGM dataset: https://github.com/Thinklab- SJTU/IMCPT-SparseGM-dataset ; Code: https://github. com/Thinklab-SJTU/ThinkMatch . Figure 1. Illustrative comparison of injective graph matching (left) and partial graph matching (right) on our remade and re- leased benchmark: IMC-PT-SparseGM (originated from IMC- PT [18], see Sec. 4for details). Green for correct matches, red for wrong matches. Without partial matching, the GM solver greed- ily matches all nodes, leading to inferior accuracy. This paper presents a general learning method to mitigate this issue. 49]. Existing GM methods mainly follow the optimiza- tion formulation by maximizing the node-wise and edge- wise afﬁnities of the matched elements, yielding an NP-hard problem known as Quadratic Assignment Problem [ 22]. The ubiquitous challenge of partial matching, how- ever, is less addressed by the existing afﬁnity-maximization pipeline. Partial matching means only a partial set of the nodes are matched, due to the existence of outliers on both sides. As shown in Fig. 1, this is commonly encountered in (visual) graph matching [ 13,16,52,53], where the exis- tence of outliers is usually unavoidable due to the errors in keypoint detectors and (self-)occlusion of objects. The recent line of deep graph matching papers [ 11,45, 52] sheds light on the partial GM, whereby the higher ca- pacity offered by deep neural networks on both the feature stage [ 52] and the solver stage [ 27,45] will hopefully dis- tinguish the outliers from inliers. However, there still lacks a general, principled approach that could enable the par- tial matching capability of existing graph matching neural networks. Some straightforward treatments such as thresh- olding [ 30], and adding dummy rows and columns [ 17] are relatively inﬂexible because their threshold and dummy val- ues are manually assigned. We aim at developing a general, uniﬁed partial match- ing handling approach, which can be readily integrated into SOTA GM networks. However, it is still an open ques- tion about how to determine whether or not a matching pair should be discarded. Besides, directly discarding the unwanted matching pairs causes non-differentiability and This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 6272 6HF 6HF 01HWZRUN &11 $WWHQWLRQIXVHG$JJUHJDWLRQZLWK,QGLYLGXDOUDSK0RGHOLQJ$)$,0RGXOH$)$80RGXOH$WWHQWLRQIXVHG$JJUHJDWLRQZLWK8QLILHG%LSDUWLWHUDSK0RGHOLQJ7RS0$OJRULWKP6HF LPDJHSDLUV GRXEO\VWRFKDVWLFPDWUL[ JUDSKZLWKIHDWXUHVIHDWXUHH[WUDFWRUJUDSKPDWFKLQJQHWZRUNPDWFKLQJUHVXOW $)$8$)$,UHILQHGPDWFKLQJPDWUL[2XU3URSRVHG0RGXOHVÁDWWHQGLIIHUHQWLDEOH62)77RSUHVKDSH Figure 2. Deep learning paradigm for partial graph matching. The input images are sent to CNN to extract features and build the input graphs. The GM network predicts a doubly-stochastic matrix based on graphs with features, followed by a differentiable top- kalgorithm. The number of inliers kis estimated by our proposed attention-fused aggregation networks. truncated gradient. We resort to the doubly-stochastic ma- trix available in most SOTA GM pipelines [ 17,44,45], whereby the values in the doubly-stochastic matrix could be viewed as the matching conﬁdence. The matchings with the top-kconﬁdence values should be preserved, assuming that the number of inliers kis known. To this end, we present a top-kdeep graph matching approach to address the partial matching challenge, whereby the differentiable top- kfor- mulation [ 48] is followed to enable the end-to-end training. Another challenge that naturally arises is how to estimate the value of k(i.e. number of inliers) from scratch. We identify the connection between the k-estimation problem and the graph-level similarity learning problem, whereby some prior successful models [ 1,2] could be further ex- ploited. In this paper, we present two networks to ef- ﬁciently reuse the mid-layers that are available in most deep GM pipelines, namely Attention- Fused Aggregation (AFA) modules, based on the similarity of graphs aggre- gated from either individual graph features (Sec. 3.2.1 ) or the doubly-stochastic similarities on a uniﬁed bipartite graph (Sec. 3.2.2 ). The AFA modules are further integrated into the learning pipeline. Besides, the severe partial matching issue that usually exists in downstream vision tasks is less addressed by exist- ing graph matching evaluation protocols. We are thus moti- vated to collect and relabel a new benchmark, namely IMC- PT-SparseGM, which is originated from the stereo matching task in Image Matching Challenge PhotoTourism (IMC-PT) 2020 [ 18]. As summarized in Tbl. 1, the new benchmark ad- dresses the severe partial matching challenge, and its graphs are of larger sizes than existing benchmarks. The main contributions of the paper are as follows. 1) We formulate the partial (graph) matching problem as a top- kscheme, i.e. in the presence of outliers in bothTable 1. Visual GM datasets. “partial rate” means the mean per- centage of occluded keypoints w.r.t. the universe of all keypoints. dataset name # visual graphs avg # nodes # universe partial rate Willow Object Class [ 6] 404 10 10 0.0% Pascal VOC Keypoint [ 5] 8702 9.07 6 to 23 28.5% IMC-PT-SparseGM-50 (ours) 25765 21.36 50 57.3% IMC-PT-SparseGM-100 (ours) 25765 44.48 100 55.5% graphs. The scheme can be integrated into SOTA GM neu- ral networks e.g. [ 17,45] as a plugin. 2) Based on the top- kformulation, we devise an end-to- end and outlier-aware neural pipeline for partial GM learn- ing and show its effectiveness. In contrast, we show that di- rectly combining the top- kscheme with either a traditional solver like RRWM [ 7] or an outlier-blind neural solver e.g. NGMv2 [ 45] still produce poor results (see Tbl. 6), sug- gesting the necessity of our integrated learning paradigm. 3) To estimate the number of inliers kwhich is often un- known in practice, we devise an attention-based supervised graph neural network whose input consists of similarity in- formation available in GM networks. With this estimation module, we enable a fully automatic neural solver for partial GM which to our best knowledge is new in the literature. 4) Our approach outperforms existing methods on pop- ular GM benchmarks notably in the setting of partial GM. Last but not least, we remake and release a new benchmark to advance the research in GM by introducing larger graphs for matching and more partial matching instances.
Wang_LipFormer_High-Fidelity_and_Generalizable_Talking_Face_Generation_With_a_Pre-Learned_CVPR_2023	Abstract Generating a talking face video from the input audio sequence is a practical yet challenging task. Most existing methods either fail to capture fine facial details or need to train a specific model for each identity. We argue that a codebook pre-learned on high-quality face images can serve as a useful prior that facilitates high-fidelity and generalizable talking head synthesis. Thanks to the strong capability of the codebook in representing face textures, we simplify the talking face generation task as finding proper lip-codes to characterize the variation of lips during portrait talking. To this end, we propose LipFormer , a Transformer-based framework to model the audio-visual coherence and predict the lip-codes sequence based on input audio features. We further introduce an adaptive face warping module, which helps warp the reference face to the target pose in the feature space, to alleviate the difficulty of lip-code prediction under different poses. By this means, LipFormer can make better use of pre- learned priors in images and is robust to posture change. Extensive experiments show that LipFormer can producemore realistic talking face videos compared to previous methods and faithfully generalize to unseen identities.	1. Introduction As an ongoing research topic, talking face generation aims to build a cross-modal mapping from an audio se- quence to a face video while maintaining natural speaking styles and audio-visual coherence. It has received growing attention in recent years due to its potential in digital humans, film-making, virtual video conferences, and online education [3, 10, 37–39, 44, 45, 47, 48, 50, 53]. A high-fidelity and generalizable talking face generation model relies heavily on high-quality (HQ) video data with vast identities. However, the existing datasets still suffer from two limitations: (1) low resolution and qualities, e.g., LRW [5] and LRS2 [1], leading to the learned model an unsatisfying synthesis quality; (2) a limited number of identities despite the clear videos, e.g., Obama [17, 31] and privately recorded data [28], which requires training a specific model for each person and it is hard to generalize to unseen portraits. These two drawbacks limit their practical applications, and it is also a challenge to collect This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 13844 a mass of such high-quality videos because they should simultaneously meet the phoneme balance and audio-visual synchronization demands. In contrast, we notice that there are many publicly available datasets of high-resolution face images, e.g., the FFHQ [15] dataset contains 70,000 identities with 1024×1024 resolutions. It helps raise a question: could these image datasets benefit the generation of a talking portrait? Fortunately, the answer is a big yes. In this work, we confirm that a high-quality pre-learned facial codebook can serve as a strong prior and facilitate talking head synthesis from the perspectives of visual fidelity and generalizability. The codebook, which is learned with the objective to reconstruct 2D faces, is capable of representing diverse face details, and hence takes most of the responsibilities to synthesize the appearance of the talking face. That way, the only thing left is to characterize the variation of lips when people talk [24, 25].We can therefore reformulate the task of talking face generation as a simpler problem, namely, finding proper lip-codes for the input face. To this end, we propose LipFormer , a Transformer- based framework for high-fidelity talking face synthesis. In particular, LipFormer first encodes the target face us- ing the pre-learned codebook, and then replaces the lip- region features with a new sequence of lip-codes that are predicted by aligning with the reference audio. Before lip- code prediction, we introduce an Adaptive Face Warping Module, which helps warp the reference face to the target pose in the feature ( i.e., codebook) space, to alleviate the texture mismatch problem caused by different poses. Last but not least, LipFormer is trained in an end-to-end way to make different modules collaborate fully, boosting the final performance. Experiments on multiple datasets confirm the superiority of our approach over state-of-the-art methods [24,51] from the perspectives of both video quality and generalizability to unseen identities.
Wang_Sharpness-Aware_Gradient_Matching_for_Domain_Generalization_CVPR_2023	Abstract The goal of domain generalization (DG) is to enhance the generalization capability of the model learned from a source domain to other unseen domains. The recently devel- oped Sharpness-Aware Minimization (SAM) method aims to achieve this goal by minimizing the sharpness measure of the loss landscape. Though SAM and its variants have demonstrated impressive DG performance, they may not al- ways converge to the desired ﬂat region with a small loss value. In this paper, we present two conditions to ensure that the model could converge to a ﬂat minimum with a small loss, and present an algorithm, named Sharpness- Aware Gradient Matching (SAGM), to meet the two condi- tions for improving model generalization capability. Specif- ically, the optimization objective of SAGM will simultane- ously minimize the empirical risk, the perturbed loss (i.e., the maximum loss within a neighborhood in the parameter space), and the gap between them. By implicitly aligning the gradient directions between the empirical risk and the perturbed loss, SAGM improves the generalization capabil- ity over SAM and its variants without increasing the com- putational cost. Extensive experimental results show that our proposed SAGM method consistently outperforms the state-of-the-art methods on ﬁve DG benchmarks, including PACS, VLCS, OfﬁceHome, TerraIncognita, and DomainNet. Codes are available at https://github.com/Wang- pengfei/SAGM .	1. Introduction Deep learning methods have achieved remarkable suc- cess in various computer vision tasks when the source data and target data are independently and identically distributed (i.i.d). However, the performance of deep models trained in a source domain can drop signiﬁcantly when applied to unseen target domains. Domain generalization (DG) aims *Corresponding authorto train a model from a set of source data such that it can be well generalized to new domains without retrain- ing. Over the past decade, research on DG has led to a plethora of methods, including those based on domain alignment [35, 16, 32, 2, 52], meta-learning [29, 3, 11, 51], and data augmentation [53, 43, 7]. Though numerous DG approaches have been proposed, a recent study called Do- mainBed [18] reveals that under a fair evaluation proto- col, the naive empirical risk minimization (ERM) method can even outperform most existing DG methods. Unfortu- nately, simply minimizing the empirical loss on a complex and non-convex loss landscape is typically insufﬁcient to achieve a good generalization capability [23, 17, 22, 15]. As a result, ERM tends to overﬁt the training data set and converge to sharp local minima [15]. Recent studies such as Sharpness-Aware Minimization (SAM) [15] try to improve the model generalization ability by minimizing the sharpness measure of the loss landscape. Denote byL()the loss to be minimized ( e.g., the cross- entropy loss for classiﬁcation), where is the parameters of the neural network. SAM ﬁrst adversarially computes a weight perturbation that maximizes the empirical risk L()and then minimizes the loss of the perturbed network, i.e.,L(+). Speciﬁcally, the objective of SAM is to mini- mize the maximum loss around the model parameter . Due to the high complexity of this min-max optimization prob- lem, SAM chooses to minimize an approximation of L(), denoted byLp()(perturbed loss, see Section 3 for details). However, minimizing Lp()is not guaranteed to converge to ﬂat minimum regions [55]. WhileLp()may not characterize well the sharpness of loss surface, the surrogate gap h(),Lp()L() can better describe it. Intuitively, the loss surface will be- come ﬂatter when h()is closer to zero because the pertur- bation parameters around will have a similar loss value. Zhuang et al. [55] proved that h()is an equivalent measure of the dominant eigenvalue of Hessian (which is the mea- sure of sharpness) at a local minimum. Therefore, we can seek ﬂat minima for better generalization ability by mini- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 3769 mizing the surrogate gap h(). Unlike SAM which only optimizes the perturbation loss Lp(), GSAM [55] jointly minimizesLp()and the surrogate gap h(). However, GSAM minimizes h()by increasing the loss L(), which will reduce the generalization performance. Zhang et al . [55] have shown that when the perturbation amplitude is sufﬁciently small, the surrogate gap is always non-negative, i.e.,Lp()L();8. Therefore, increasing L()will also increase the difﬁculty of optimizing Lp(;D). We propose two conditions that should be met to obtain a model with good generalization performance. (i) First, the loss within a neighborhood of the desired minimum should be sufﬁciently low. (ii) Second, the minimum is within a ﬂat loss surface. More speciﬁcally, condition (i) implies a low training loss and represents good performance on the source training data, while condition (ii) reduces the performance gap of the model on training and testing data. Based on the above analysis, we propose a new DG method, namely Sharpness-Aware Gradient Matching (SAGM), to simultaneously minimize three objectives, the empirical riskL(), the perturbed loss Lp()and the surro- gate gaph(). By minimizingL()andLp(), we search for a region with low loss, which satisﬁes condition (i). By minimizingh(), we avoid steep valleys, which meets con- dition (ii). However, optimizing these three objectives si- multaneously is difﬁcult due to the inevitable gradient con- ﬂicts during training. Fortunately, we ﬁnd that when the gradient directions of L()andLp()are consistent, the gradient direction of h()is also consistent with them, and hence the gradient descent can be effectively applied to all the three losses. Therefore, we transform the optimiza- tion objective into minimizing L(),Lp(), and the angle between their gradients, and achieve this goal by implic- itly aligning the gradient directions between the L()and Lp(). The proposed SAGM improves the generalization performance of the model by facilitating the model to con- verge to a ﬂat region with a small loss value. Compared with SAM, our SAGM does not increase the computational cost. In addition, SAGM can be combined with previous data augmentation methods, such as Mixstyle [54] for fur- ther performance improvements. The contributions of this work are summarized as fol- lows. First, we analyze the limitations of SAM-like meth- ods and propose two conditions to ensure the model conver- gence to a ﬂat region with a small loss. Second, we propose the SAGM algorithm to improve the DG capability of deep models. Finally, we demonstrate the superior performance of SAGM to state-of-the-arts on ﬁve DG benchmarks.
Xie_High-Fidelity_3D_GAN_Inversion_by_Pseudo-Multi-View_Optimization_CVPR_2023	Abstract We present a high-fidelity 3D generative adversarial net- work (GAN) inversion framework that can synthesize photo- realistic novel views while preserving specific details of the input image. High-fidelity 3D GAN inversion is inherently challenging due to the geometry-texture trade-off, where overfitting to a single view input image often damages the estimated geometry during the latent optimization. To solve this challenge, we propose a novel pipeline that builds on the pseudo-multi-view estimation with visibility analysis. We keep the original textures for the visible parts and uti- lize generative priors for the occluded parts. Extensive ex- periments show that our approach achieves advantageous reconstruction and novel view synthesis quality over prior work, even for images with out-of-distribution textures. The proposed pipeline also enables image attribute editing with the inverted latent code and 3D-aware texture modifica- tion. Our approach enables high-fidelity 3D rendering from a single image, which is promising for various applica- tions of AI-generated 3D content. The source code is at https://github.com/jiaxinxie97/HFGI3D/ .	1. Introduction Real-world 3D-aware editing with a single 2D image fas- cinates various essential applications in computer graphics, such as virtual reality ( VR), augmented reality ( AR), and immersive meetings. Recent advancement in 3D GANs [12, 13,21,45] has achieved photo-realistic 3D-consistent image generation. With the GAN inversion approaches [19,51,72], which can map the images to the latent space of the pre- trained 3D-aware model, high-fidelity 3D-aware editing be- comes promising. High-fidelity 3D-aware inversion aims to generate novel views with high-quality reconstruction and 3D consistency, but existing methods can hardly meet these two goals si- multaneously. Although current GAN inversion methods based on 2D GANs [1, 53, 54, 69] can perform 3D-related attributes (e.g., head pose) editing, the generated view is *Joint first authors †Joint corresponding authorsinconsistent due to the lack of the underlying 3D represen- tation. When applying the existing optimization-based in- version approaches on 3D-aware GANs [12,36], we can re- trieve high-fidelity reconstruction by overfitting to a single input image. However, different from 2D GAN inversion, the reconstruction quality of 3D GAN depends not only on the input view’s faithfulness but also on the quality of the synthesized novel views. During the optimization process, the obvious artifacts in the synthesized novel views occur with the appearance of high-fidelity details in the input view as analyzed in Sec. 3. As only a single image is available in the optimization process, the reconstruction suffers from extreme ambiguity: infinite combinations of color and den- sity can reconstruct the single input image, especially with out-of-distribution textures. Based on the above observation, we propose our 3D- aware inversion pipeline by optimizing the reconstruction not only on the input image but also on a set of pseudo- multi-views. The pseudo views provide additional regular- ization, and thus the ambiguity is greatly reduced. Estimat- ing the pseudo views is non-trivial as it requires maintaining the texture details while also generating the occluded parts in a plausible manner, based on the input view. We first esti- mate an initial geometry and conduct a visibility analysis to solve these challenges. We directly utilize the textures from the input image for the visible parts to preserve the texture details. For the occluded parts, we use a pretrained gen- erator to synthesize the reasonable inpainted regions. With the additional supervision from the pseudo-multi-views, our approach achieves high-fidelity reconstruction results with the correct 3D geometry. Our approach enables two types of editing: latent at- tributes editing and 3D-aware texture modification. We fol- low the previous work [55] and calculate the attribute direc- tion in the latent code space. By modifying the inverted la- tent code in a specific direction, we can control the general attribute (e.g., smile, ages for portraits as in Figure 1(b)). Since the proposed pipeline enables inversion with out-of- distribution textures, we can achieve compelling 3D con- sistent editing by only modifying the textures of the input images (e.g., stylization or adding a tattoo in Figure 1(c)). In summary, we propose a high-fidelity 3D GAN in- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 321 Input image Reconstruction Novel view 1 Novel view 2 Novel view 3 (a) High-fidelity 3D GAN inversionSmile+ Age+ (b) Latent Attributes Editing (c) 3D-aware Textures Modification Figure 1. High-fidelity 3D GAN inversion results on real-world images with two types of editing ability. Our method preserves compelling details and achieves high 3D consistency. version method by pseudo-multi-view optimization given an input image. Our approach can synthesize compelling 3D-consistent novel views that are visually and geometri- cally consistent with the input image. We perform exten- sive quantitative and qualitative experiments, which demon- strate that our 3D GAN inversion approach outperforms other 2D/3D GAN inversion baselines in both photorealism and faithfulness.
Wang_PET-NeuS_Positional_Encoding_Tri-Planes_for_Neural_Surfaces_CVPR_2023	Abstract A signed distance function (SDF) parametrized by an MLP is a common ingredient of neural surface reconstruc- tion. We build on the successful recent method NeuS to ex- tend it by three new components. The first component is to borrow the tri-plane representation from EG3D and rep- resent signed distance fields as a mixture of tri-planes and MLPs instead of representing it with MLPs only. Using tri- planes leads to a more expressive data structure but will also introduce noise in the reconstructed surface. The sec- ond component is to use a new type of positional encoding with learnable weights to combat noise in the reconstruc- tion process. We divide the features in the tri-plane into multiple frequency scales and modulate them with sin and cos functions of different frequencies. The third component is to use learnable convolution operations on the tri-plane features using self-attention convolution to produce features with different frequency bands. The experiments show that PET-NeuS achieves high-fidelity surface reconstruction on standard datasets. Following previous work and using the Chamfer metric as the most important way to measure sur- face reconstruction quality, we are able to improve upon the NeuS baseline by 57% on Nerf-synthetic (0.84 compared to 1.97) and by 15.5% on DTU (0.71 compared to 0.84). The qualitative evaluation reveals how our method can better control the interference of high-frequency noise.	1. Introduction Implicit neural functions, or neural fields, have received a lot of attention in recent research. The seminal paper NeRF [25] combines neural fields with volume rendering, enabling high-quality novel view synthesis. Inspired by NeRF, NeuS [41] and V olSDF [44] introduce a signed dis- tance function (SDF) into the volume rendering equation and regularize the SDF, so that smooth surface models can be reconstructed. However, these methods use pure MLP networks to encode SDFs. Although these two methods can reconstruct smooth surfaces, they both leave room for im- provement when it comes to reconstructing surface details. One research direction ( [5, 6, 26, 33, 46]) explores datastructures such as tri-planes or voxel grids that are suitable to improve the NeRF framework, in terms of speed or recon- struction quality. However, data structures that are success- ful for novel view synthesis may not bring immediate suc- cess when employed for surface reconstruction as shown in the third column of Fig. 1. While a greater expressiveness to encode local details is useful to better fit the input data, there is also less inductive bias towards a smooth surface. There- fore, noise during image acquisition, high-frequency shad- ing, or high-frequency texture variations are more likely to result in a noisy reconstructed surface. In our work, we explore how to increase expressiveness to encode local features while at the same time reducing the impact of noise interference. We choose to build on the tri- plane data structure since it consumes less memory and can be easier scaled to higher resolutions. In our work, we build on EG3D and NeuS to propose a novel framework, called PET-NeuS. First, we propose a method to integrate the tri-plane data structure into a sur- face reconstruction framework in order to be able to model an SDF with more local details. Second, since the features between tri-plane pixels do not share learnable parameters, we use positional encoding to modulate the tri-plane fea- tures, thereby enhancing the smoothness of the learnable features. Third, the positional encoding involves functions of different frequencies. In order to better match differ- ent frequencies, we propose to use multi-scale self-attention convolution kernels with different window sizes to perform convolution in the spatial domain to generate features of dif- ferent frequency bands. This further increases the fidelity of the surface reconstruction while suppressing noise. We experiment on two datasets to verify the effectiveness of our method, the DTU dataset and the NeRF-Synthetic dataset. Since the DTU dataset contains non-Lambertian surfaces, the ability of the network to resist noise interfer- ence can be verified. The NeRF-Synthetic dataset has many sharp features, which can verify that our framework can ef- fectively utilize its improved local expressiveness to better reconstruct local details. We show superior performance compared to state-of-the-art methods on both datasets. In summary, our contributions are as follows: • We propose to train neural implicit surfaces with a tri- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 12598 Figure 1. The challenge of using the tri-plane representation directly. First column: reference image. Second to the fifth column: NeuS, Learning SDF using tri-planes, OURS without self-attention convolution, and OURS. plane architecture to enable the reconstructed surfaces to better preserve fine-grained local features. • We derive a novel positional encoding strategy to be used in conjunction with tri-plane features in order to reduce noise interference. • We utilize self-attention convolution to produce tri- plane features with different frequency bands to match the positional encoding of different frequencies, fur- ther improving the fidelity of surface reconstruction.
Wei_iCLIP_Bridging_Image_Classification_and_Contrastive_Language-Image_Pre-Training_for_Visual_CVPR_2023	Abstract This paper presents a method that effectively com- bines two prevalent visual recognition methods, i.e., image classification and contrastive language-image pre-training, dubbed iCLIP . Instead of na ¨ıve multi-task learning that use two separate heads for each task, we fuse the two tasks in a deep fashion that adapts the image classification to share the same formula and the same model weights with the language-image pre-training. To further bridge these two tasks, we propose to enhance the category names in image classification tasks using external knowledge, such as their descriptions in dictionaries. Extensive experiments show that the proposed method combines the advantages of two tasks well: the strong discrimination ability in im- age classification tasks due to the clean category labels, and the good zero-shot ability in CLIP tasks ascribed to the richer semantics in the text descriptions. In particu- lar, it reaches 82.9% top-1 accuracy on IN-1K, and mean- while surpasses CLIP by 1.8%, with similar model size, on zero-shot recognition of Kornblith 12-dataset benchmark. The code and models are publicly available at https: //github.com/weiyx16/iCLIP .	1. Introduction Image classification is a classic visual problem whose goal is to classify images into a fixed set of pre-defined cat- egories. For example, the widely used ImageNet dataset [8] carefully annotated 14 million images and categorize them into 21,841 categories chosen from the WordNet [36]. For image classification, each category provides a clear taxon- omy that groups images of the same category together and separates images from different categories, and thus endows the learnt representation with strong discriminant ability. However, this classification ability is limited to a fixed set of categories [8, 29, 51]. *Corresponding Author. The work is done when Yixuan Wei, Zhuliang Yao, and Zhenda Xie are interns at Microsoft Research Asia. Alt-texts Dictionary enhanced classesText Encoder Visual Encoder Contrastive Loss Classes Dictionary night bird a photo of a night bird , any bird associated with night, like owl... Dictionary Idx: 585 Figure 1. An illustration of the proposed iCLIP framework. The iCLIP framework can take two types of annotations for training: classes and alt-texts. It converts the conventional image classifi- cation formula to share the same text encoder and the same co- sine classifier as that used in the contrastive language-image pre- training (CLIP). It also uses a dictionary-enhanced approach to enrich the original class names in the image classification prob- lem with external information involved in dictionaries. The deep fusion and knowledge-enriched classes both greatly improve the performance compared to na ¨ıve multi-task learning or performing one of the two tasks alone. Recently, the method that learns to contrast image-text pairs, known as contrastive language-image pre-training (abbr. CLIP), has well made up such shortage of the con- ventional image classification methods to achieve strong zero-shot recognition ability [24, 44]. These methods em- ploy a contrastive learning framework, where images and their corresponding alt-texts are treated as positive pairs, while images with all other alt-texts are treated as negative pairs. Thanks to the rich semantics involved in the alt-texts, the images can be weakly connected to almost arbitrary cat- egories that already appear in the alt-texts, resulting in its zero-shot ability. A drawback is that the image-text pairs are usually crawled from the internet without human label- ing, leading to their noisy and ambiguous nature. Thus the learnt representations are often not conceptual compact, and may lack certain discriminative ability. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 2776 This paper explores how to effectively combine these two powerful visual recognition and representation learn- ing methods, to take advantages of both methods and data sources while relieving their shortages. We first try a na ¨ıve multi-task learning framework that applies the original head networks of the two tasks on top of a shared visual encoder, and jointly learn the network with separate losses of the two tasks. This na ¨ıve multi-task learning approach has been able to benefit each individual tasks, but the effect is marginal. We thus seek to fuse the two tasks more deeply, so that the advantages of the two tasks can be more effectively joined for better visual recognition, as well as for better transfer- able representations. To this end, our first technique is to deeply unify the formulations of image classification and CLIP learning. By examining their formulations, we found there are two main differences: 1) Different classification losses. Image classification tasks typically use a linear classification loss which has better fitting ability due to the non-normalized nature, while the CLIP-based methods adopt a cosine clas- sifier which has better transferability for new domains and categories [2, 6, 9, 18, 38, 57]. 2) Different parameteriza- tion methods for classifier weights. Image classification tasks usually directly optimize the parametric classification weights without a need to process text semantics in class names. The CLIP method can be regarded as generating classifier weights through a text encoder and learns the text encoder instead. The text-encoder-based classifier allows sharing between alt-texts as well as modeling their relation- ships, which enables the ability to tackle any classes. Although the linear classifier and direct classifier weight parameterization have been common practice in image clas- sification for many years, it is interesting to find that chang- ing the old formulation as that in the CLIP approach has almost no performance degradation for pure image classifi- cation problems. This indicates that we can directly adapt the image classification formulation to the cosine classifier and the text encoder parameterization used by CLIP, with almost no loss. This also allows us to further share the text encoder for both class names and alt-texts. Our experiments show that this deep fusion approach performs much better than the na ¨ıve multi-task method for both in-domain/zero- shot classification and multi-modal retrieval tasks learning (see 3). Another gap between the image classification and CLIP lies in the different text richness. Class names are usually in short, i.e., one or a few words, and sometimes are even ambiguous and polysemous in referring to specific seman- tics, for example, “ night bird ” can represents either “ owl” or “nightingale ”. On the contrary, alt-texts in CLIP are usu- ally full sentences containing rich information. To further bridge the gap between the image classification and CLIP, we propose a second technique that leverages the knowledgebase to enhance the original class names, such as the expla- nations in dictionaries. In our implementation, knowledge is simply encoded as a prefix/suffix prompt, as illustrated in Fig 1. Although simple, dictionary enhanced method shows to maintain the accuracy for pure image classification prob- lem (see Table 1), while greatly improve the zero-shot and multi-modal retrieval performance as shown in Table 2 and 3. Note the process is just like human beings who learn new words or concepts through both real examples and explana- tions in dictionaries. By these techniques, we present a framework that deeply fuses the two important tasks of image classification and contrastive language-image pre-training, dubbed iCLIP. Ex- tensive experiments using different combinations of image classification and image-text pair datasets show that the iCLIP method can take advantages of both the discrimina- tive power of image classification tasks and the zero-shot ability in CLIP-like tasks, and perform significantly bet- ter than conducting each task alone or the na ¨ıve multi-task learning in both the in-domain/zero-shot classification and multi-modal retrieval problems. The iCLIP method also shows that learning a stronger transferable representation than using each of the two tasks alone, verified on a vari- ety of downstream tasks, including ADE20K semantic seg- mentation [68], LVIS long-tail detection [17], and video ac- tion recognition [26], as well as different evaluation settings of few-shot and fine-tuning. Our contributions are summa- rized as follows: • We combined two important vision tasks of im- age classification and contrastive language-image pre- training into a single framework. • We found that the original image classification for- mulation can be adapted to CLIP approach with al- most no performance degradation. With this finding, we present a deep fusion approach in which the two tasks share the same text encoder and the same clas- sifier type, whose effectiveness is extensively verified on benchmarks. • We proposed a simple yet effective method to intro- duce knowledge bases into image classification, ad- dressing the ambiguous and polysemous issue of the originally short image names as well as further bridges the gap between classes and alt-texts. It also provides the first showcase of applying knowledge bases into computer vision problems.
Wei_LEGO-Net_Learning_Regular_Rearrangements_of_Objects_in_Rooms_CVPR_2023	Abstract Humans universally dislike the task of cleaning up a messy room. If machines were to help us with this task, they must understand human criteria for regular arrange- ments, such as several types of symmetry, co-linearity or co-circularity, spacing uniformity in linear or circu- lar patterns, and further inter-object relationships that re- late to style and functionality. Previous approaches for this task relied on human input to explicitly specify goal state, or synthesized scenes from scratch – but such meth- ods do not address the rearrangement of existing messy scenes without providing a goal state. In this paper, we present LEGO-Net, a data-driven transformer-based it- erative method for LEarning re Gular rearrangement of Objects in messy rooms. LEGO-Net is partly inspired by diffusion models – it starts with an initial messy state and iteratively “de-noises” the position and orientation of ob- jects to a regular state while reducing distance traveled. Given randomly perturbed object positions and orientations ∗Core contribution.in an existing dataset of professionally-arranged scenes, our method is trained to recover a regular re-arrangement. Results demonstrate that our method is able to reliably re- arrange room scenes and outperform other methods. We additionally propose a metric for evaluating regularity in room arrangements using number-theoretic machinery.	1. Introduction What makes the arrangement of furniture and objects in a room appear regular? While exact preferences may vary, humans have by-and-large universally shared criteria of reg- ular room arrangements: for instance, heavy cabinets are arranged to align with walls, chairs are positioned evenly around a table in linear or circular configurations, or night stands are placed symmetrically on the two sides of a bed. Humans also share a common dislike of physically perform- ing the task of rearranging a messy room. To build auto- mated robotic systems that can guide or actually rearrange objects in a room, we first need methods that understand the shared human criteria for regular room rearrangements and respect the physical constraints of rearrangements. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19037 Human criteria for regular rearrangements can be sub- tle and complex, including geometric rules of reflexional, translational, or rotational symmetry, linear or circular alignments, and spacing uniformity. Functional and stylistic inter-object relationships are also important: for example, a TV tends to be in front of and facing a sofa, chairs are next to a table, etc. Many of these criteria interact and, at times, conflict with one another. As a result, in general, there is more than one desirable clean arrangement for any given messy arrangement. In our setting, we further desire that the clean rearrangement we create to be informed by the initial messy arrangement – and not be entirely different – for multiple reasons. First, there may have been a particular clean arrangement that gave rise to the messy one – and it may be desirable to recover a similar arrangement. Second, we want to minimize the motion of objects as much as pos- sible to respect the physical constraints and effort involved – especially the motion of big and heavy furniture. Unfortu- nately, extant methods fail to capture these criteria: methods for scene synthesis from scratch [25, 29, 37, 69, 70, 72] ig- nore the initial state of objects in a room, and rearrangement methods often require scene-specific human input in the form of a goal state [1, 46] or language description [27, 47]. In this paper, we present LEGO-Net, a method for LEarning re Gular rearrangement of Objects in rooms di- rectly from data. Different from work that focuses on ar- ranging new objects from scratch or requires goal state specification, we focus on rearranging existing objects without any additional input at inference time. We take as input the position, orientation, class label, and extents of room objects in a specific arrangement, and output a room with the same objects but regularly re-arranged. LEGO-Net uses a transformer-based architecture [53] that is, in part, motivated by recent denoising diffusion probabilistic mod- els that learn a reverse diffusion process for generative mod- eling [16, 49, 50]. We learn human criteria for regular rear- rangements from a dataset of professionally designed clean (regular) scenes [15], and represent each scene as a collec- tion of objects and a floor plan. Prior to training, we perturb the regular scenes to generate noisy configurations. During training, our transformer learns to predict the original, de- noised arrangement from the perturbed scene and its floor plan. During inference, instead of directly re-arranging scenes with our model, which would amount to na ¨ıve re- gression, we run a Langevin dynamics-like reverse process to iteratively denoise object positions and orientations. This iterative process retains the flavor of original room state, while limiting object movement during re-arrangement. We conduct extensive experiments on public datasets to show that our approach realistically rearranges noisy scene arrangements, while respecting initial object positions. We also demonstrate that our method is able to generalize to previously unseen collection of objects in a wide variety offloor plans. Furthermore, we include extensive experimen- tal results (e.g., Fig. 1 and Fig. 4), including a new metric to evaluate regularity of re-arrangements, aimed at measuring the presence of sparse linear integer relationships among object positions in the final state (using the PSLQ algo- rithm [13]). To sum up, we contribute: • A generalizable, data-driven method that learns to reg- ularly re-arrange the position and orientation of ob- jects in various kinds of messy rooms. • An iterative approach to re-arrangement at inference time that retains flavor of the original arrangement and minimizes object travel distance. • An in-depth analysis of the performance and charac- teristics of the denoising-based scene rearrangement. • A new metric to measure the regularity of object ar- rangements based on integer relation algorithms.
Wang_Few-Shot_Learning_With_Visual_Distribution_Calibration_and_Cross-Modal_Distribution_Alignment_CVPR_2023	Abstract Pre-trained vision-language models have inspired much research on few-shot learning. However, with only a few training images, there exist two crucial problems: (1) the visual feature distributions are easily distracted by class-irrelevant information in images, and (2) the alignment between the visual and language feature distri- butions is difficult. To deal with the distraction problem, we propose a Selective Attack module, which consists of trainable adapters that generate spatial attention maps of images to guide the attacks on class-irrelevant image areas. By messing up these areas, the critical features are captured and the visual distributions of image features are calibrated. To better align the visual and language feature distributions that describe the same object class, we pro- pose a cross-modal distribution alignment module, in which we introduce a vision-language prototype for each class to align the distributions, and adopt the Earth Mover’s Distance (EMD) to optimize the prototypes. For efficient computation, the upper bound of EMD is derived. In addition, we propose an augmentation strategy to increase the diversity of the images and the text prompts, which can reduce overfitting to the few-shot training images. Extensive experiments on 11 datasets demonstrate that our method consistently outperforms prior arts in few-shot learning. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv /SADA.	1. Introduction Thanks to the availability of large-scale datasets and well-designed training strategies, the performances of many computer vision tasks have been greatly improved. Re- cent progress in vision-language models (VLMs), such as *Co-first author. †Corresponding author.CLIP [29] and ALIGN [17], provides a promising way towards utilizing human language to address downstream recognition tasks efficiently. As vision and language usually contain complementary information, joint learning of image and text representations has proven quite effective. Al- though CLIP has demonstrated impressive zero-shot learn- ing capability, it is still challenging to better adapt it to downstream tasks. Naively fine-tuning CLIP on down- stream datasets has limited effect, since it may destroy the prior learned from the massive data during pre-training. Therefore, effective transfer methods are needed to boost the downstream performances of CLIP. In order to main- tain the capability of pre-trained VLMs and further boost downstream performances, different approaches have been proposed to fine-tune a small proportion of additional pa- rameters while keeping the pre-trained parameters frozen. Among these approaches, prompt learning [42, 43] and vi- sual adapters [13, 41] are two common approaches. How- ever, the lack of training samples in few-shot settings increases the risk of overfitting the trained prompts or adapters. The class-irrelevant features ( e.g., the cluttered image backgrounds) drive the image features far away from their true distributions of the same category. Besides, VLMs such as CLIP have such a problem that the distributions of the image and text features are not really aligned [30], and the problem becomes more challenging in few-shot settings. Therefore, the visual distributions should be calibrated by reducing class-irrelevant image contents, and the distribu- tions of image and text features should be further aligned, so as to promote the model’s learning of class-relevant crit- ical features. The purpose of this paper is to develop an ef- fective VLM transfer strategy for few-shot learning to solve the above problems with Selective Attack (SA) and Cross- Modal Distribution Alignment (CMDA). Images often contain class-irrelevant information, which is also embedded into the image representations. With only a few samples, the model can easily learn these cluttered representations, resulting in overfitting. This seriously hin- ders the learning of critical features that help the model rec- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 23445 EncodingEncodingSelectiveAttackCross-ModalDistributionAlignmentHistogram of features(a)(b)(c)(d) ClassClassClassClass…p1pMp1p1p1……pMpMpMLearnable promptsFigure 1. (a) The t-SNE [35] visualization of the image feature distribution before Selective Attack, where the features are obtained by the CLIP image encoder on the CIFAR10 dataset. The dots in different colors represent different classes of the image features. (b) After Selective Attack, the intra-class distribution is significantly more compact. (c) The distribution histograms of image features and text features of the same class (‘bird’) on CIFAR10 before CMDA, where the horizontal axis denotes the value of each element of the feature vectors, and the vertical axis denotes the number of elements. (d) After CMDA, the difference between the two distributions is significantly reduced. ognize unseen samples. To solve this problem, we propose the SA module, which consists of two trainable adapters that generate a kernelized attention map to locate the class- irrelevant areas of the images. The attention is adopted to guide Gaussian perturbations to attack images before they are fed into the image encoder. By messing up these class- irrelevant image contents through SA, we facilitate the model’s learning of truly critical features that can be trans- ferred to recognize new samples within the same category. As an example in Figs. 1 (a) and (b), after Selective Attack (SA), the distributions of the image features are calibrated, and the intra-class features become obviously more clus- tered. Another challenge is that the distributions of the image and the text representations of the same class are not truly aligned in CLIP [30] as shown in Fig. 1(c). The unaligned distributions lead to inaccurate similarity calculations be- tween image features and text features during inference, re- sulting in incorrect predictions. The lack of samples in few- shot settings further makes the problem even more serious. To address it, we propose a CMDA module, in which we construct a Vision-Language Prototype (VLP) for each class to promote the cross-modal distribution alignment. Specif- ically, the element values of VLP are initialized by aver- aging all the image representations from the corresponding class. During training, each VLP is optimized by reducing its distance to the language prototype (defined in Sec. 3.4) of the same class, thus promoting the cross-modal distri- bution alignment. The Earth Mover’s Distance (EMD) is a suitable metric for the alignment, which can not only reflect the similarity between two distributions but also represent the minimal transmission cost [40]. We derive a concise upper bound of the EMD distance, which can balance theperformance and computational consumption. As shown in Figs. 1 (c) and (d), the effect of Cross-Modal Distribu- tion Alignment (CMDA) is obvious that the difference be- tween the image and text feature distributions is effectively reduced. In this way, the image features after CMDA can be better predicted by the text features. Automatic prompt learning for pre-trained VLMs has been proposed to reduce the expensive cost of hand-crafted prompt engineering [43]. However, the learned prompts may suffer from more overfitting than manual prompts [42]. Therefore, instead of learning one soft prompt, we learn a distribution over a collection of prompts, as in ProDA [23]. Moreover, we introduce an augmentation strategy to in- crease the diversity of the images and the prompts. Specifi- cally, we search for the four best augmentations from a col- lection of predefined ones. Using these operations, each image is augmented into four different forms. The collec- tion of prompts is also divided into four groups, with each group trained by images in the corresponding augmentation form. Through the strategy, we improve the diversity of the images and the prompts, and fully excavate the semantic in- formation in the prompts. The framework of our method is shown in Fig. 2. Our contributions are summarized as follows: • We conduct Selective Attack on the class-irrelevant regions of images with the guidance of the attention generated by two trainable adapters to facilitate the model’s learning of class-related features, which cal- ibrates the visual distributions. • We propose Cross-Modal Distribution Alignment op- timized by an EMD loss. The upper bound of EMD for Gaussian distribution is further derived for compu- tation efficiency. 23446 +TextEncoderImageEncoderImagefeatureTextfeatureSelective AttackEMD(image,text)TrainingImageEncoder+InferenceImagefeatureTextfeatureClass 1Class 2Class n… ∗+ TextEncoderAssemble …1 2 nsimilarity∗𝑆JCMDA𝑆 𝑆VLPs VLPsClassClassClassClassJ…p1pMp1p1p1……pMpMpM ClassClassClassClassJ…p1pMp1p1p1……pMpMpMTrainable promptsFigure 2. Overview of our framework. We introduce a Selective Attack module to reduce the intra-class distances of image features during training. We also design a Cross-Modal Distribution Alignment (CMDA) module to align the distributions of image and text representations. During training, the trainable parameters are denoted in orange and the encoders of CLIP are frozen. J: the number of augmentations; Ⓢ: cosine similarity computation; ⊛: element-wise product. • We present an augmentation strategy to reduce overfit- ting and increase the diversity of images and prompts. • Our method outperforms prior arts in few-shot learning on 11 benchmarks.
Weng_3D_Human_Keypoints_Estimation_From_Point_Clouds_in_the_Wild_CVPR_2023	Abstract Training a 3D human keypoint detector from point clouds in a supervised manner requires large volumes of high quality labels. While it is relatively easy to capture large amounts of human point clouds, annotating 3D key- points is expensive, subjective, error prone and especially difficult for long-tail cases (pedestrians with rare poses, scooterists, etc.). In this work, we propose GC-KPL - Geometry Consistency inspired Key Point Leaning, an ap- proach for learning 3D human joint locations from point clouds without human labels. We achieve this by our novel unsupervised loss formulations that account for the struc- ture and movement of the human body. We show that by training on a large training set from Waymo Open Dataset [21] without any human annotated keypoints, we are able to achieve reasonable performance as compared to the fully supervised approach. Further, the backbone benefits from the unsupervised training and is useful in downstream few- shot learning of keypoints, where fine-tuning on only 10 per- cent of the labeled training data gives comparable perfor- mance to fine-tuning on the entire set. We demonstrated that GC-KPL outperforms by a large margin over SoTA when trained on entire dataset and efficiently leverages large vol- umes of unlabeled data.	1. Introduction Estimation of human pose in 3D is an important prob- lem in computer vision and it has a wide range of appli- cations including AR/VR, AI-assisted healthcare, and au- tonomous driving [4, 29, 32]. For autonomous systems, be- ing able to perceive human poses from sensor data ( e.g. Li- DAR point clouds) is particularly essential to reason about the surrounding environment and make safe maneuvers. Despite the high level of interest in human pose estima- tion in the wild, only few papers approached outdoor 3D keypoint detection using point cloud. A main reason is that *Work done as an intern at Waymo. Transformer Encoder Keypoint locations learned with Unsupervised Losses Downstream Fine-tuning Small amount of labeled point clouds Pre-trained Backbone MLP In-the-wild Point Cloud 3D keypoints Figure 1. We present GC-KPL, a novel method for learning 3D human keypoints from in-the-wild point clouds without any human labels. We propose to learn keypoint locations using unsupervised losses that account for the structure and movement of the human body. The backbone learns useful semantics from unsupervised learning and can be used in down- stream fine-tuning tasks to boost the performance of 3D keypoint estima- tion. training a pedestrian pose estimation model requires large amount of high quality in-the-wild data with ground truth labels. Annotating 3D human keypoints on point cloud data is expensive, time consuming and error prone. Although there are a few existing point cloud datasets with ground truth human poses [11, 13, 21], they are limited in terms of the quantity of the 3D annotations and diversity of the data. Therefore, fully-supervised human keypoint detectors trained on such datasets do not generalize well for long tail cases. For this reason, previous approaches on pedestrian 3D keypoint estimation have mainly focused on utilizing 2D weak supervision [4, 32] which is easier to obtain, or lever- aging signals from others modalities ( e.g. RGB, depth) [29]. Nonetheless, there is a lot of useful information in the large amount of unlabeled LiDAR data that previous works on human pose estimation have not made an effort to utilize. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1158 In this work, we propose a novel and effective method for learning 3D human keypoints from in-the-wild point clouds without using any manual labeled 3D keypoints. Our ap- proach is built on top of the key observation that human skeletons are roughly centered within approximately rigid body parts and that the location and movement of the sur- face points should explain the movement of the skeleton and vice versa. To that end, we design novel unsupervised loss terms for learning locations of the 3D keypoints/skeleton within human point clouds which correspond to 3D loca- tions of major joints of human body. In the proposed method, we first train a transformer- based regression model for predicting keypoints and a se- mantic segmentation model for localizing body parts on a synthetic data constructed from randomly posed SMPL hu- man body model [15]. Then, we train on the entire Waymo Open Dataset [21] without using any 3D ground-truth anno- tation of human keypoints. Through unsupervised training, keypoint predictions are refined and the backbone learns useful information from large amount of unannotated data. In summary, we make the following contributions: • We present GC-KPL, a method for learning human 3D keypoints for in-the-wild point clouds without any manual keypoint annotations. • Drawing insight from the structure and movement of the human body, we propose three effective and novel unsupervised losses for refining keypoints. We show that the proposed losses are effective for unsupervised keypoint learning on Waymo Open Dataset. • Through downstream fine-tuning/few-shot experi- ments, we demonstrate that GC-KPL can be used as unsupervised representation learning for human point clouds, which opens up the possibility to utilize a prac- tically infinite amounts of sensor data to improve hu- man pose understanding in autonomous driving.
Xu_Learning_Imbalanced_Data_With_Vision_Transformers_CVPR_2023	Abstract The real-world data tends to be heavily imbalanced and severely skew the data-driven deep neural networks, which makes Long-Tailed Recognition (LTR) a massive challeng- ing task. Existing LTR methods seldom train Vision Trans- formers (ViTs) with Long-Tailed (LT) data, while the off- the-shelf pretrain weight of ViTs always leads to unfair comparisons. In this paper, we systematically investigate the ViTs’ performance in LTR and propose LiVT to train ViTs from scratch only with LT data. With the observa- tion that ViTs suffer more severe LTR problems, we con- duct Masked Generative Pretraining (MGP) to learn gener- alized features. With ample and solid evidence, we show that MGP is more robust than supervised manners. Al- though Binary Cross Entropy (BCE) loss performs well with ViTs, it struggles on the LTR tasks. We further propose the balanced BCE to ameliorate it with strong theoreti- cal groundings. Specially, we derive the unbiased exten- sion of Sigmoid and compensate extra logit margins for de- ploying it. Our Bal-BCE contributes to the quick conver- gence of ViTs in just a few epochs. Extensive experiments demonstrate that with MGP and Bal-BCE, LiVT success- fully trains ViTs well without any additional data and out- performs comparable state-of-the-art methods significantly, e.g., our ViT-B achieves 81.0% Top-1 accuracy in iNatural- ist 2018 without bells and whistles. Code is available at https://github.com/XuZhengzhuo/LiVT .	1. Introduction With the vast success in the computer vision field, Vision Transformers (ViTs) [15, 43] get increasingly popular and have been widely used in visual recognition [15], detec- tion [5], and video analysis [16]. These models are heavily dependent on large-scale and balanced data to avoid overfit- ting [39,52,82]. However, real-world data usually confronts severe class-imbalance problems, i.e., most labels (tail) are associated with limited instances while a few categories (head) occupy dominant samples. The models simply clas- sify images into head classes for lower error because the 0 20 40 60 80 100 280 300 320303540455055606570Top-1 Accuracy (%) Parameters (M)Baseline NCL RIDE DeiT MAE LiVT(Ours) LiVT/384(Ours)ViT-Tiny ViT-Small ViT-Base ViT-Large R50RIDE-4E R50NCL R50×3[13] [34] [61] [55] [18] Figure 1. Top-1 Acc v.s. Model Size on ImageNet-LT dataset. We choose the Tiny / Small / Base / Large ViT and multi-expert approaches. R50 represents the ResNet50 model. ViT-Base gets lower Acc than ResNet50 when trained in a supervised manner. head always overwhelms tail ones in LTR. The data paucity also results in the model overfitting on the tail with unac- cepted generalization. The aforementioned problems make LongTailRecognization (LTR) a challenging task. Numerous papers [4,13,22,34,35,44,70] handle the LTR problem with traditional supervised cross-entropy learning based on ResNet [20] or its derivatives [68]. Some meth- ods use ViTs with pretrained weights on ImageNet [52] (or larger datasets), which leads to unfair comparisons with ad- ditional data, e.g.on ImageNet-LT (a subset of ImageNet- 1K) benchmark. Moreover, there are still limited explo- rations on the utilization of Long-Tailed (LT) data to train ViTs effectively. Therefore, in this paper, we try to train ViTs from scratch with LT data . We observe that it is par- ticularly difficult to train ViT with LT labels’ supervision. As Tab. 1 shows, ViTs degrade heavily when training data become skewed. ViT-B is much worse than ResNet50 with the same CE training manner ( c.f. Fig. 1). One reason- able explanation is that ViTs require longer training to learn the inductive bias, while CNNs offer the built-in translation invariance implicitly. Yet another one lies in the label sta- tistical bias in the LTR datasets, which confuses models to make predictions with an inherent bias to the head [12, 47]. The well-trained ViTs have to overcome the above plights simultaneously to avoid falling into dilemmas. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 15793 Inspired by decoupling [29], many methods [9, 12, 60, 80, 83] attempt to enhance feature extraction in supervised manners like mixup [74] / remix [9], or Self-Supervised Learning (SSL) like Contrastive Learning (CL) [7,19]. Liu et al. [41] claim that SSL representations are more robust to class imbalance than supervised ones, which inspires us to train ViTs with SSL. However, CL is quite challenging for extensive memory requisition and converge difficulties [8], where more explorations are required to work well with ViTs in LTR. In contrast, we propose to Learn imbalanced data with ViTs (LiVT) by Masked Generative Pretraining (MGP) and Balanced FineTuning (BFT). Firstly, LiVT adopts MGP to enhance ViTs’ feature ex- traction, which has been proven effective on BeiT [2] and MAE [18]. It reconstructs the masked region of images with an extra lightweight decoder. We observe that MGP is stable with ViTs and robust enough to LT data with em- pirical evidence. Despite the label distribution, the compa- rable number of training images will bring similar feature extraction ability, which greatly alleviates the toxic effect of LT labels [26]. Meanwhile, the training is accelerated by masked tokens with acceptable memory requisition. Secondly, LiVT trains the downstream head with rebal- ancing strategies to utilize annotation information, which is consistent with [29, 35, 80]. Generally, Binary Cross- Entropy (BCE) loss performs better than Cross-Entropy loss when collaborating with ViTs [55]. However, it fails to catch up with widely adopted Balanced Cross- Entropy (Bal-CE) loss and shows severe training instability in LTR. We propose the Balanced BCE (Bal-BCE) loss to revise the mismatch margins given by Bal-CE. Detailed and solid the- oretical derivations are provided from Bayesian theory. Our Bal-BCE ameliorates BCE by a large margin and achieves state-of-the-art (SOTA) performance with ViTs. Extensive experiments show that LiVT learns LT data more efficiently and outperforms vanilla ViT [15], DeiT III [55], and MAE [18] remarkably. As detailed compar- isons in Fig. 1, LiVT achieves SOTA on ImageNet-LT with affordable parameters, despite that ImageNet-LT is a rel- atively small dataset for ViTs. The ViT-Small [55] also achieves outstanding performance compared to ResNet50. Our key contributions are summarized as follows. • To our best knowledge, we are the first to investigate training ViTs from scratch with LT data systematically. • We pinpoint that the masked generative pretraining is robust to LT data, which avoids the toxic influence of imbalanced labels on feature learning. • With a solid theoretical grounding, we propose the bal- anced version of BCE loss (Bal-BCE), which improves the vanilla BCE by a large margin in LTR. • We propose LiVT recipe to train ViTs from scratch, and the performance of LiVT achieves state-of-the-art across various benchmarks for long-tailed recognition.Table 1. Top-1 accuracy (%) of different recipes to train ViT-B-16 from scratch on ImageNet-LT/BAL. All perform much worse on LT than BAL. See descriptions of LT & BAL in section 5.1. Dataset ViT ∆ DeiT III ∆ MAE ∆ ImageNet-BAL 38.7 - 67.2 - 69.2 - ImageNet-LT 31.6 -7.0 48.4 -18.8 54.5 -14.7
Xu_EqMotion_Equivariant_Multi-Agent_Motion_Prediction_With_Invariant_Interaction_Reasoning_CVPR_2023	Abstract Learning to predict agent motions with relationship rea- soning is important for many applications. In motion pre- diction tasks, maintaining motion equivariance under Eu- clidean geometric transformations and invariance of agent interaction is a critical and fundamental principle. However, such equivariance and invariance properties are overlooked by most existing methods. To fill this gap, we propose Eq- Motion, an efficient equivariant motion prediction model with invariant interaction reasoning. To achieve motion equivariance, we propose an equivariant geometric feature learning module to learn a Euclidean transformable feature through dedicated designs of equivariant operations. To reason agent’s interactions, we propose an invariant interac- tion reasoning module to achieve a more stable interaction modeling. To further promote more comprehensive motion features, we propose an invariant pattern feature learning module to learn an invariant pattern feature, which coop- erates with the equivariant geometric feature to enhance network expressiveness. We conduct experiments for the proposed model on four distinct scenarios: particle dynam- ics, molecule dynamics, human skeleton motion prediction and pedestrian trajectory prediction. Experimental results show that our method is not only generally applicable, but also achieves state-of-the-art prediction performances on all the four tasks, improving by 24.0/30.1/8.6/9.2%. Code is available at https://github.com/MediaBrain- SJTU/EqMotion .	1. Introduction Motion prediction aims to predict future trajectories of multiple interacting agents given their historical observations. It is widely studied in many applications like physics [3, 28], molecule dynamics [7], autonomous driving [35] and human- robot interaction [38,68]. In the task of motion prediction, an *Corresponding author. (a) Particle dynamics (b) Molecule dynamics (c) Human skeleton motion(d) Pedestrian trajectories SpringPast motionFuture motion(equivariant)StickInteraction(invariant) Single bondPast motionFuture motion (equivariant)Double bondInteraction(invariant) Bone connectionPast motionFuture motion(equivariant)Interaction(invariant)Past motionFuture motion (equivariant)MeetingInteraction(invariant)Euclidean transformationEuclidean transformation EuclideantransformationEuclideantransformation Figure 1. Motion equivariance and interaction invariance under the Euclidean geometric transformation is a fundamental principle for a prediction model, but this principle is often overlooked by previous works. In this work, we propose EqMotion to fill this gap. often-overlooked yet fundamental principle is that a predic- tion model is required to be equivariant under the Euclidean geometric transformation (including translation, rotation and reflection), and at the same time maintain the interaction relationships invariant. Motion equivariance here means that if an input motion is transformed under a Euclidean trans- formation, the output motion must be equally transformed under the same transformation. Interaction invariance means that the way agents interact remains unchanged under the input’s transformation. Figure 1 shows real-world examples of motion equivariance and interaction invariance. Employing this principle in a network design brings at least two benefits. First, the network will be robust to arbi- trary Euclidean transformations. Second, the network will have the capability of being generalizable over rotations and translations of the data. This capability makes the network more compact, reducing the network’s learning burden and contributing to a more accurate prediction. Despite the motion equivariance property being important This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1410 and fundamental, it is often neglected and not guaranteed by most existing motion prediction methods. The main reason is that these methods transform the input motion sequence directly into abstract feature vectors, where the geometric transformations are not traceable, causing the geometric relationships between agents to be irretrievable. Random augmentation will ease the equivariance problem, but it is still unable to guarantee the equivariance property. [30] uses non-parametric pre and post coordinate processing to achieve equivariance, but its parametric network structures do not satisfy equivariance. Some methods propose equivariant parametric network structures utilizing the higher-order rep- resentations of spherical harmonics [14, 61] or proposing an equivariant message passing [58], but they focus on the state-to-state prediction. This means that they use only one historical timestamp to predict one future timestamp. Conse- quently, these methods have limitations on utilizing motion’s temporal information and modeling interaction relationships since a single-state observation is insufficient for both inter- action modeling and temporal dependency modeling. In this paper, we propose EqMotion, the first motion pre- diction model that is theoretically equivariant to the input motion under Euclidean geometric transformations based on the parametric network. The proposed EqMotion has three novel designs: equivariant geometric feature learning, invariant pattern feature learning and invariant interaction reasoning. To ensure motion equivariance, we propose an equivariant geometric feature learning module to learn a Eu- clidean transformable geometric feature through dedicated designs of equivariant operations. The geometric feature preserves motion attributes that are relevant to Euclidean transformations. To promote more comprehensive repre- sentation power, we introduce an invariant pattern feature learning module to complement the network with motion attributes that are independent of Euclidean transformations. The pattern features, cooperated into the geometric features, provide expressive motion representations by exploiting mo- tions’ spatial-temporal dependencies. To further infer the interactions during motion prediction, we propose an invariant interaction reasoning module, which ensures that the captured interaction relationships are invari- ant to the input motion under Euclidean transformations. The module infers an invariant interaction graph by utilizing invariant factors in motions. The edge weights in the inter- action graph categorize agents’ interactions into different types, leading to better interaction representation. We conduct extensive experiments on four different sce- narios to evaluate our method’s effectiveness: particle dy- namics, molecule dynamics, 3D human skeleton motion and pedestrian trajectories. Comparing to many task-specific motion prediction methods, our method is generally appli- cable and achieves state-of-the-art performance in all these tasks by reducing the prediction error by 24.0/30.1/8.6/9.2 %respectively. We also present that EqMotion is lightweight, and has a model size less than 30 %of many other models’ sizes. We show that EqMotion using only 5 %data can achieve a comparable performance with other methods that take full data. As a summary, here are our contributions: •We propose EqMotion, the first motion prediction model that theoretically ensures sequence-to-sequence mo- tion equivariance based on the parametric network. With equivariance, EqMotion promotes more generalization abil- ity of motion feature learning, leading to more robust and accurate prediction. •We propose a novel invariant interaction reasoning mod- ule, in which the captured interactions between agents are invariant to the input motion under Euclidean geometric transformations. With this, EqMotion achieves more gener- alization ability and stability in the interaction reasoning. •We conduct experiments on four types of scenarios and find that EqMotion is applicable to all these different tasks, and importantly outperforms existing state-of-the-art methods on all the tasks.
Wang_Open-Set_Fine-Grained_Retrieval_via_Prompting_Vision-Language_Evaluator_CVPR_2023	Abstract Open-set fine-grained retrieval is an emerging challenge that requires an extra capability to retrieve unknown sub- categories during evaluation. However, current works focus on close-set visual concepts, where all the subcategories are pre-defined, and make it hard to capture discrimina- tive knowledge from unknown subcategories, consequently failing to handle unknown subcategories in open-world sce- narios. In this work, we propose a novel Prompting vision- Language Evaluator (PLEor) framework based on the re- cently introduced contrastive language-image pretraining (CLIP) model, for open-set fine-grained retrieval. PLEor could leverage pre-trained CLIP model to infer the discrep- ancies encompassing both pre-defined and unknown subcat- egories, called category-specific discrepancies, and trans- fer them to the backbone network trained in the close-set scenarios. To make pre-trained CLIP model sensitive to category-specific discrepancies, we design a dual prompt scheme to learn a vision prompt specifying the category- specific discrepancies, and turn random vectors with cate- gory names in a text prompt into category-specific discrep- ancy descriptions. Moreover, a vision-language evaluator is proposed to semantically align the vision and text prompts based on CLIP model, and reinforce each other. In addi- tion, we propose an open-set knowledge transfer to transfer the category-specific discrepancies into the backbone net- work using knowledge distillation mechanism. Quantitative and qualitative experiments show that our PLEor achieves promising performance on open-set fine-grained datasets.	1. Introduction Open-set fine-grained retrieval (OSFR) attempts to build a well-generalized embedding space where the visual dis- crepancies among unknown subcategories are clearly re- flected. It plays a vital role in numerous vision applica- *Corresponding author: hjli@dlut.edu.cn ℱ𝐶𝑁𝑁 0.15 0.02 0.94 ⋮ 0.230 0 1 ⋮ 0ℒ𝐶𝐸 𝑪𝑳𝑨𝑺𝑺𝑨(a) a classification-based evaluator (e.g., [34, 43, 52, 64]) ℱ𝐶𝑁𝑁 𝑪𝑳𝑨𝑺𝑺𝑨 𝑪𝑳𝑨𝑺𝑺𝑩ℒ𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒 𝑪𝑳𝑨𝑺𝑺𝑨(b) a metric-based evaluator (e.g., [3, 20, 39, 46, 51, 63]) 𝑪𝑳𝑨𝑺𝑺𝑨ℱ𝐶𝑁𝑁 𝐶𝐿𝐼𝑃 +𝑪𝑳𝑨𝑺𝑺𝑨 ℒ𝐾𝑛𝑜𝑤𝑙𝑒𝑑𝑔𝑒 𝑇𝑟𝑎𝑛𝑠𝑓𝑒𝑟 𝑫𝒐𝒈 𝑪𝒂𝒕⋯ 𝑪𝒂𝒓 𝑭𝒐𝒙 𝑷𝒊𝒈 𝑻𝒐𝒚⋯ 𝑷𝒐𝒕 𝑨𝒊𝒓Open -world Knowledge A brown dog is sitting on the grass. Two striped cats are lying on bed. Tibetan fox walks on the grasslandOpen -world Knowledge Tuned Frozen 𝑪𝑳𝑨𝑺𝑺𝑨 Vison Prompt (Visual discrepancies ) Text Prompt (Category -specific descriptions) (c) a vision-language evaluator with open-world knowledge Figure 1. Comparison on existing evaluators in open-set fine- grained retrieval. Although our PLEor (c) is trained in a close- set scenarios, similar with previous works (a) (b), it could mine the category-specific discrepancies using pre-trained CLIP model aided by vision and text prompts, and transfer the discrepancies encompassing both pre-defined and unknown subcategories to our model. This enables our model to procure in-depth understanding for unknown subcategories owing to distilling the knowledge with open-world visual concepts from CLIP model, improving retrieval performance eventually in open-set scenarios. tions from fashion industry, e.g., retrieval of diverse types of clothes [1, 31], to environmental conservation, e.g., retriev- ing endangered species [7,49,50]. As shown in Fig. 1(a)(b), existing works follow a close-set learning setting, where all the subcategories are pre-defined, and evaluate embeddings identifying the visually similar objects of pre-defined sub- categories. However, such evaluation focuses on closed-set visual concepts, limiting the model to a pre-defined list of subcategories, and is not generalizable when it comes to un- known subcategories unseen during training. Fortunately, recent works [66, 67] using large-scale con- trastive language-image pretraining (CLIP) model [37] have shown great potentials in alleviating this limitation. As shown in Fig. 1(c), CLIP model is pretrained from scratch on a dataset of 400 moillion image-text pairs, which are au- tomatically collected from the publicly available sources on This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 19381 the Internet. Based on this, CLIP model could associate much wider range of visual concepts in the images with their text descriptions, rather than a fixed set of pre-defined categories. Therefore, one question naturally arises: is it possible that we can effectively exploit the open-set visual concepts in CLIP model to solve OSFR task? It is already answered yes by recent studies exploring how to transfer the knowledge from CLIP model to other downstream tasks via prompt techniques [10, 16, 26, 37, 56, 66, 67]. However, their prompt strategies are tailored for capturing category- level semantic ( e.g., dog and cat) rather than more detailed visual discrepancies for distinguishing fine-grained objects (e.g., different breeds of dogs). Therefore, how to effec- tively make pre-trained CLIP model sensitive to the visual discrepancies encompassing both pre-defined and unknown subcategories (termed as category-specific discrepancies), and transfer these discrepancies to the model trained in closed-set scenarios is worthy of investigation. To this end, we design a novel prompting vision- language evaluator (PLEor) for OSFR, based on the power of recently introduced CLIP model. Technically, to make pre-trained CLIP model sensitive to category-specific dis- crepancy, we design a dual prompt scheme composed of vision prompt and text prompt for explicitly highlighting the category-specific discrepancies from the input perspec- tive. Concretely, the vision prompt specifies the category- specific discrepancies via parsing semantic features inferred by the backbone network. And the text prompt turns ran- dom vectors with category names into category-specific dis- crepancy descriptions. Meanwhile, a vision-language eval- uator is proposed to encourage pre-trained CLIP model to locate the category-specific descriptions in vision prompt and generate the category-specific visual semantics into text prompt. In this way, the OSFR task aided by the designed prompts is close to the solved task of pre-training CLIP model, thus making the CLIP model sensitive to category- specific discrepancy. Nevertheless, a non-negligible prob- lem is that the corporation of CLIP model and backbone network is quite complex, leading to very time consuming and memory demanding during evaluation. Thereby, we propose an open-set knowledge transfer module to transfer the category-specific discrepancies from CLIP model to the backbone network using knowledge distillation mechanism. Our contributions are summarized as follows: • A prompting vision-language evaluator, i.e., PLEor, is proposed. It can distill the knowledge with open-world visual concepts from CLIP model to alleviate the prob- lems behind open-set scenarios. To our best knowl- edge, we are the first to regard CLIP model as an eval- uator specifically for OSFR task. • PLEor provides timely insights into the adaptation of pre-trained CLIP model adopting prompt learning, andcrucially, demonstrates the effectiveness of a simple modification for inputs of CLIP model in OSFR. • PLEor achieves new state-of-the-art results compared with classification-based and metric-based evaluators, which is significant gains of 8.0% average retrieval ac- curacy on three widely-used OSFR datasets.
Wen_Learnable_Skeleton-Aware_3D_Point_Cloud_Sampling_CVPR_2023	Abstract Point cloud sampling is crucial for efficient large-scale point cloud analysis, where learning-to-sample methods have recently received increasing attention from the com- munity for jointly training with downstream tasks. However, the above-mentioned task-specific sampling methods usu- ally fail to explore the geometries of objects in an explicit manner. In this paper, we introduce a new skeleton-aware learning-to-sample method by learning object skeletons as the prior knowledge to preserve the object geometry and topology information during sampling. Specifically, with- out labor-intensive annotations per object category, we first learn category-agnostic object skeletons via the medial axis transform definition in an unsupervised manner. With ob- ject skeleton, we then evaluate the histogram of the local feature size as the prior knowledge to formulate skeleton- aware sampling from a probabilistic perspective. Addition- ally, the proposed skeleton-aware sampling pipeline with the task network is thus end-to-end trainable by explor- ing the reparameterization trick. Extensive experiments on three popular downstream tasks, point cloud classification, retrieval, and reconstruction, demonstrate the effectiveness of the proposed method for efficient point cloud analysis.	1. Introduction With the development of 3D sensing technologies, ac- quiring 3D data becomes more accessible than before, and there are a growing number of data repositories available online, such as ModelNet [62], ShapeNet [6], ScanNet [11], and KITTI [18]. Among popular 3D shape representa- tions such as point cloud [41], voxel [72], mesh [55], and multi-view images [51], the point cloud is becoming in- creasingly popular as the first-hand data captured by Li- DAR or depth camera (e.g., Kinect), which has been widely applied in various applications such as scene reconstruc- tion [19, 26], autonomous driving navigation [35], and vir- tual reality (VR) [59]. Though a high-resolution point cloud *Equal contribution (a) (b) (c) (d) Figure 1. (a) Point cloud; (b) Object skeleton; (c) Random sam- pling; (d) Skeleton-aware sampling. Compared with random sam- pling, skeleton-aware sampling tends to preserve the object geom- etry and topology information. can accurately capture the geometry and topology details of complex objects, it remains challenging for those devices with limited computation and storage resources. Therefore, point cloud sampling, aiming to find a small set of points to represent the object shape and topology effectively, is usu- ally indispensable for efficient large-scale point cloud anal- ysis [24, 29, 32, 33, 41, 42, 57, 61, 71]. Traditional point cloud sampling methods such as ran- dom sampling (RS) and farthest point sampling (FPS) [15, 42] usually select a subset of points directly using the raw data information [7, 42, 43, 47, 68]. Specifically, RS is very efficient but may miss sparse regions, while FPS has better coverage on the entire point set but suffers from the latency bottleneck in parallel computation. To improve the per- formances on downstream tasks, learn-to-sample methods have been recently proposed to jointly optimize the sam- pling algorithm and each specific downstream task [10, 14, 20,27,56]. Though considerable progress has been achieved in downstream tasks such as point cloud classification and reconstruction, one critical issue remains poorly investi- gated: as objects usually have complex topology structures and irregular surface morphology, it is challenging to pre- serve the object’s geometrical information during the point cloud sampling process. Skeleton is an efficient representation to capture the un- derlying object shape structures, which has been widely used as the structural abstraction in many visual understand- ing tasks [31, 37, 53]. Inspired by this, we build our point cloud sampling strategy on top of the object skeleton to This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 17671 preserve different objects’ intrinsic geometrical and topo- logical natures. Here we illustrate a comparison between random sampling and skeleton-aware sampling in Fig. 1. However, the skeleton extraction is usually non-trivial due to the following reasons [30, 52]: 1) given the diversity of object topological structures, it is difficult to have a con- sistent category-agnostic skeleton definition at the semantic level, where existing datasets usually consider only single or a few known object categories, such as human skele- ton; 2) topological methods are usually category-agnostic by emphasizing geometrical and topological properties of the shape, such as its connectivity, topology, length, direc- tion, and width. Nevertheless, they usually require a sub- stantial amount of time for geometrical processing and are also notoriously sensitive to surface noise. Therefore, we resort to the medial axis transform (MAT) definition of ob- ject skeleton and deep neural networks to learn effective skeletal representations in an unsupervised manner. With the learned object skeleton, we then formulate the skeleton-aware point cloud sampling pipeline from a proba- bilistic perspective. Specifically, we first calculate the local feature size (LFS) [2] for each point to measure the object’s size near a particular point. We then explore the LFS distri- bution as the prior sampling probability on each point and use the LFS histogram in practice since per-point LFS is usually sensitive to point cloud noise. By learning object skeletons with deep neural networks, we have the skeleton- aware prior probability for sampling each point. To adapt skeleton-aware sampling for downstream tasks, we jointly optimize the posterior sampling probability on each point in an end-to-end manner. Notably, by exploring the categori- cal reparameterization with Gumbel-softmax [22], the cate- gorical sampling based on LFS histogram is differentiable. Therefore, both sampling and task networks are end-to-end learnable for task-specific point cloud sampling. In this pa- per, our main contributions can be summarized as follows: 1. We propose a new skeleton-aware point cloud sam- pling method to preserve the object geometry and topology information during sampling. 2. We introduce the category-agnostic skeleton learning in an unsupervised manner to provide the prior knowl- edge for skeleton-aware point cloud sampling. 3. We conduct extensive experiments on three important downstream tasks, point cloud classification, retrieval, and reconstruction, to evaluate the proposed approach.
Wang_METransformer_Radiology_Report_Generation_by_Transformer_With_Multiple_Learnable_Expert_CVPR_2023	Abstract In clinical scenarios, multi-specialist consultation could significantly benefit the diagnosis, especially for intricate cases. This inspires us to explore a “multi-expert joint diagnosis” mechanism to upgrade the existing “single ex- pert” framework commonly seen in the current literature. To this end, we propose METransformer, a method to real- ize this idea with a transformer-based backbone. The key design of our method is the introduction of multiple learn- able “expert” tokens into both the transformer encoder and decoder. In the encoder, each expert token interacts with both vision tokens and other expert tokens to learn to attend different image regions for image representation. These ex- pert tokens are encouraged to capture complementary in- formation by an orthogonal loss that minimizes their over- lap. In the decoder, each attended expert token guides the cross-attention between input words and visual tokens, thus influencing the generated report. A metrics-based expert voting strategy is further developed to generate the final re- port. By the multi-experts concept, our model enjoys the merits of an ensemble-based approach but through a man- ner that is computationally more efficient and supports more sophisticated interactions among experts. Experimental re- sults demonstrate the promising performance of our pro- posed model on two widely used benchmarks. Last but not least, the framework-level innovation makes our work ready to incorporate advances on existing “single-expert” models to further improve its performance.	1. Introduction Interpreting radiology images (e.g., chest X-ray) and writing diagnostic reports are essential operations in clinical practice and normally require a considerable manual work-load. Therefore, radiology report generation, which aims to automatically generate a free-text description based on a radiograph, is highly desired to ease the burden of radiolo- gists while maintaining the quality of health care. Recently, substantial progress has been made towards research on au- tomated radiology report generation models [4,5,15–17,21, 22,33–35,41,43,44,47]. Most existing studies adopt a con- ventional encoder-decoder architecture following the image captioning paradigm [6, 25, 32, 37, 48] and resort to opti- mizing network structure or introducing external or prior information to aid report generation. These methods, in this paper, are collectively referred to as “single-expert” based diagnostic captioning methods. However, diagnostic report generation is a very challeng- ing task as disease anomalies usually only occupy a small portion of the whole image and could appear at arbitrary lo- cations. Due to the fine-grained nature of radiology images, it is hard to focus on the correct image regions throughout the report generation procedure despite different attentions developed in recent works [16, 21]. Meanwhile, it is no- ticed that in clinic scenarios, multi-specialist consultation is especially beneficial for those intricate diagnostic cases that challenge a single specialist for a comprehensive and accu- rate diagnosis. The above observations led us to think, could we design a model to simulate the multi-specialist consul- tation scenario? Based on this motivation, we propose a new diagnostic captioning framework, METransformer, to mimic the “multi-expert joint diagnosis” process. Built upon a transformer backbone, METransformer introduces multiple “expert tokens”, representing multiple experts, into both the transformer encoder and decoder. Each expert to- ken learns to attend distinct image regions and interacts with other expert tokens to capture reliable and complementary visual information and produce a diagnosis report in paral- lel. The optimal report is selected through an expert voting strategy to produce the final report. Our design is based on This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 11558 the assumption that it would be easier for multiple experts than a single one to capture visual patterns of clinical im- portance, which is verified by our experimental results. Specifically, we feed both the expert tokens (learnable embeddings) and the visual tokens (image patches embed- dings) into the Expert Transformer encoder which is com- prised of a vision transformer (ViT) encoder and a bilinear transformer encoder. In ViT encoder, each expert token in- teracts not only with the visual tokens but also with the other expert tokens. Further, in the bilinear transformer encoder, to enable each “expert” to capture fine-grained image in- formation, we compute higher-order attention between ex- pert tokens and visual tokens, which has proved to be effec- tive in fine-grained classification tasks [20]. It is notewor- thy that the expert tokens in the encoder are encouraged to learn complementary representations by an orthogonal loss so that they attend differently to different image regions. With these carefully learned expert token embeddings, in the decoder, we take them as a guide to regulate the learn- ing of word embeddings and visual token embedding in the report generation process. This results in M different word and visual token embeddings, thus producing M candidate reports, where M is the number of experts. We further pro- pose a metric-based expert voting strategy to generate the final report from the M candidates. By utilizing multiple experts, our model, to some extent, is analogous to ensemble-based approaches, while each ex- pert token corresponds to an individual model. While it en- joys the merits of ensemble-based approaches, our model is designed in a manner that is computationally more efficient and supports more sophisticated interactions among ex- perts. Therefore, it can scale up with only a trivial increase of model parameters and achieves better performance, as demonstrated in our experimental study. Our main contributions are summarized as follows. First, we propose a new diagnostic captioning frame- work, METransformer, which is conceptually “multi-expert joint diagnosis” for radiology report generation, by intro- ducing learnable expert tokens and encouraging them to learn complementary representations using both linear and non-linear attentions. Second, our model enjoys the benefits of an ensemble approach. Thanks to the carefully designed network struc- ture and the end-to-end training manner, our model can achieve better results than common ensemble approaches while greatly reducing training parameters and improving training efficiency. Third, our approach shows promising performance on two widely used benchmarks IU-Xray and MIMIC-CXR over multiple state-of-the-art methods. The clinic relevance of the generated reports is also analyzed.
Weng_Event-Based_Blurry_Frame_Interpolation_Under_Blind_Exposure_CVPR_2023	Abstract Restoring sharp high frame-rate videos from low frame- rate blurry videos is a challenging problem. Existing blurry frame interpolation methods assume a predefined and known exposure time, which suffer from severe perfor- mance drop when applied to videos captured in the wild. In this paper, we study the problem of blurry frame interpola- tion under blind exposure with the assistance of an event camera. The high temporal resolution of the event camera is beneficial to obtain the exposure prior that is lost during the imaging process. Besides, sharp frames can be restored using event streams and blurry frames relying on the mu- tual constraint among them. Therefore, we first propose an exposure estimation strategy guided by event streams to es- timate the lost exposure prior, transforming the blind expo- sure problem well-posed. Second, we propose to model the mutual constraint with a temporal-exposure control strat- egy through iterative residual learning. Our blurry frame interpolation method achieves a distinct performance boost over existing methods on both synthetic and self-collected real-world datasets under blind exposure.	1. Introduction Blurry frame interpolation (BFI) [23, 46, 73] aims to re- store sharp high frame-rate videos from blurry low frame- rate videos, which is highly desirable for a wide range of applications, such as novel view interpolation synthesis [7], frame rate conversion [32], slow motion [21] and inter- frame video compression [64]. Compared with the cascade scheme, i.e., combining frame deblurring [20,25,33,39,40, 51, 53, 59] with frame interpolation [2, 21, 31, 32, 35, 37, 38, 68], the joint method [46,73] is more effective, which over- comes the error accumulation problem and ambiguity of the temporal scope [46]. Despite remarkable improvement, prior works [23, 46] *Corresponding author. This work was supported in part by the Na- tional Natural Science Foundation of China under Grants 62131003 and 62021001. non- blind exposure readoutvariable and unknown exposure exposureshutter period shutter periodblind exposure predefined and known + exposure estimation mutual constraint … … (a) (b) …Figure 1. (a) Non-blind exposure setting and blind exposure set- ting. (b) Illustrative flow of our solution to blurry frame interpola- tion under blind exposure. assume that the exposure time is predefined or known as the shutter period, which we call non-blind exposure as shown in the left part of Fig. 1 (a). However, the com- plicated video capturing in the wild gives rise to variable and unknown exposure time, which we call blind exposure . The right part of Fig. 1 (a) presents that the shutter period is the summation of exposure time and data readout time. For better imaging, the exposure time is variable to fit the changing light conditions in the real imaging environments, especially when the auto-exposure function turns on [73]. Challenges and Motivation. In this paper, we focus on the problem of BFI under blind exposure, which is prac- tical yet has barely been investigated explicitly. The first challenge of this problem can be attributed to the intrinsic imaging mechanism of frame-based cameras. As can be seen from Fig. 1 (a), the accumulation operation of frame- based cameras inevitably results in the loss of motion infor- mation during the exposure time. Particularly, the variable exposure time further leads to the temporal jittering, which degrades the performance of the video enhancement algo- rithms with constant/predefined exposure time assumption. Another challenge is that there misses a decent model as the effective guidance. Conventional frame-based methods for BFI are vulnerable to the artifacts introduced by flow-based warping or straight forward prediction. To overcome the above challenges, we attempt to pro- vide a new perspective, by resorting to a novel sensor. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 1588 Event cameras [8, 69], also known as neuromorphic sen- sors, are bio-inspired visual sensors that output events by detecting spatiotemporal brightness changes. We demon- strate that event cameras are suitable for BFI under blind exposure from two points as shown in Fig. 1 (b). First, the high temporal resolution property of event cameras is able to compensate for the exposure prior that is lost dur- ing the imaging process of frame-based cameras. To this end, we propose an exposure estimation strategy guided by event streams to estimate the lost exposure prior. In such a way, the blind exposure problem can be made well-posed, which eases the difficulty of video restoration. Second, the event stream is a natural constraint between blurry frames and sharp frames, providing a physical model for video restoration. To effectively model this physical correlation, we propose a temporal-exposure control strategy that takes timestamps and exposure priors as inputs for interpolation through iterative residual learning. By exploiting the pro- posed strategies of exposure estimation and time-exposure control, we are able to perform arbitrary-time interpolation from blurry frames under blind exposure. To evaluate our method in real-world scenarios, we collect a real blurry video dataset using a DA VIS-346 color event camera, which includes multiple exposure assumptions. Experimental re- sults on synthetic and self-collected real datasets demon- strate our superior performance over existing methods. The contributions of this paper can be summarized as follows: 1) we provide a decent solution for BFI under blind exposure by using event cameras, for the first time; 2) we propose an exposure estimation strategy guided by event streams, which makes the blind exposure problem well- posed; 3) we propose a temporal-exposure control strategy, which enables BFI at an arbitrary timestamp under blind exposure; 4) we achieve superior performance over existing state-of-the-art methods on both synthetic and self-collected real-world datasets.
Wang_Learning_Bottleneck_Concepts_in_Image_Classification_CVPR_2023	Abstract Interpreting and explaining the behavior of deep neural networks is critical for many tasks. Explainable AI pro-vides a way to address this challenge, mostly by provid-ing per-pixel relevance to the decision. Yet, interpretingsuch explanations may require expert knowledge. Some re-cent attempts toward interpretability adopt a concept-basedframework, giving a higher-level relationship between someconcepts and model decisions. This paper proposes Bot- tleneck Concept Learner (BotCL), which represents an im-age solely by the presence/absence of concepts learnedthrough training over the target task without explicit super-vision over the concepts. It uses self-supervision and tai-lored regularizers so that learned concepts can be human- understandable. Using some image classiﬁcation tasks as our testbed, we demonstrate BotCL’s potential to rebuildneural networks for better interpretability 1.	1. Introduction Understanding the behavior of deep neural networks (DNNs) is a major challenge in the explainable AI (XAI)community, especially for medical applications [19,38], foridentifying biases in DNNs [2, 18, 42], etc. Tremendous re- search efforts have been devoted to the post-hoc paradigm for a posteriori explanation [29, 33]. This paradigm pro- duces a relevance map to spot regions in the input image thatinteract with the model’s decision. Yet the relevance maponly tells low-level (or per-pixel) relationships and does not explicitly convey any semantics behind the decision. Inter- pretation of relevance maps may require expert knowledge. The concept-based framework [22, 37, 50] is inspired by the human capacity to learn a new concept by (subcon-sciously) ﬁnding ﬁner-grained concepts and reuse them indifferent ways for better recognition [24]. Instead of giv-ing per-pixel relevance, this framework offers higher-level Corresponding author. 1Code is avaliable at https://github.com/wbw520/BotCL and a simple demo is available at https://botcl.liangzhili.com/. OQVUJNBHF $PODFQUCPUUMFOFDLSFQSFTFOUBUJPO %JTDPWFSFEDPODFQUT%BUBTFUCpt.1 Cpt.2 Cpt.3 Cpt.4 Cpt.5 Cpt.6 Figure 1. Examples of concepts discovered by BotCL in Ima- geNet [10] and concepts in the input image. BotCL automatically discovers a set of concepts optimized for the target task and repre-sents an image solely with the presence/absence of concepts. relationships between the image and decision mediated by a limited number of concepts . That is, the decision is ex- plained by giving a set of concepts found in the image. Theinterpretation of the decision is thus straightforward oncethe interpretation of each concept is established. Some works use concepts for the post-hoc paradigm for better interpretation of the decision [14, 50], while the linkbetween the decision and concepts in the image is not ob-vious. The concept bottleneck structure [23] uses the pres- ence/absence of concepts as image representation (referredto as concept activation ). The classiﬁer has access only to the concept activation, so the decision is strongly tied to the concepts. This bottleneck structure has become the main-stream of the concept-based framework [5, 20, 28, 31]. A major difﬁculty in this framework is designing a set of concepts that suits the target task. A promising approach ishandcrafting them [4, 21, 48], which inherently offers bet-ter interpretability at the cost of extra annotations on theconcepts. Recent attempts automatically discover concepts[1, 13, 14, 46]. Such concepts may not always be consis-tent with how humans (or models) see the world [25, 47]and may require some effort to interpret them, but conceptdiscovery without supervision is a signiﬁcant advantage. This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 10962 Inspired by these works, we propose bottleneck c oncept learner (BotCL) for simultaneously discovering concepts and learning the classiﬁer. BotCL optimizes concepts forthe given target image classiﬁcation task without supervi-sion for the concepts. An image is represented solely by the existence of concepts and is classiﬁed using them. Weadopt a slot attention-based mechanism [26, 27] to spot the region in which each concept is found. This gives an extrasignal for interpreting the decision since one can easily seewhat each learned concept represents by collectively show-ing training images with the detected concepts. Figure 1 shows examples from ImageNet [10]. BotCL discovers apredeﬁned number of concepts in the dataset, which are ex- empliﬁed by several images with attention maps. An imageof Great White Shark is represented by the right part of mouth ( Cpt.1 ) and ﬁns ( Cpt.3 ). BotCL uses a single fully-connected (FC) layer as a classiﬁer, which is simplebut enough to encode the co-occurrence of each concept andeach class. Contribution . For better concept discovery, we propose to use self-supervision over concepts , inspired by the re- cent success in representation learning [9, 16]. Our ablationstudy demonstrates that self-supervision by contrastive lossis the key. We also try several constraints on concepts them-selves, i.e.,individual consistency to make a concept more selective and mutual distinctiveness for better coverage of various visual elements. These additional constraints regu-lar the training process and help the model learn conceptsof higher quality.
Wang_Decoupling-and-Aggregating_for_Image_Exposure_Correction_CVPR_2023	Abstract The images captured under improper exposure condi- tions often suffer from contrast degradation and detail dis- tortion. Contrast degradation will destroy the statisti- cal properties of low-frequency components, while detail distortion will disturb the structural properties of high- frequency components, leading to the low-frequency and high-frequency components being mixed and inseparable. This will limit the statistical and structural modeling ca- pacity for exposure correction. To address this issue, this paper proposes to decouple the contrast enhancement and detail restoration within each convolution process. It is based on the observation that, in the local regions covered by convolution kernels, the feature response of low-/high- frequency can be decoupled by addition/difference opera- tion. To this end, we inject the addition/difference operation into the convolution process and devise a Contrast Aware (CA) unit and a Detail Aware (DA) unit to facilitate the statistical and structural regularities modeling. The pro- posed CA and DA can be plugged into existing CNN-based exposure correction networks to substitute the Traditional Convolution (TConv) to improve the performance. Fur- thermore, to maintain the computational costs of the net- work without changing, we aggregate two units into a single TConv kernel using structural re-parameterization. Evalu- ations of nine methods and five benchmark datasets demon- strate that our proposed method can comprehensively im- prove the performance of existing methods without intro- ducing extra computational costs compared with the origi- nal networks. The codes will be publicly available.	1. Introduction Images captured under improper exposure conditions of- ten suffer from under-exposure or over-exposure problems [2, 14, 16]. Improper exposure will change the statistical distribution of image brightness, resulting in contrast degra- *Co-first author.†Corresponding author. Figure 1. (a, b) The PSNR and SSIM comparison of TConv and our DAConv on the ME dataset. (c, d) The PSNR and SSIM com- parison of TConv and our DAConv on the SICE dataset. Under the boosting of DAConv, the performance of existing methods has been comprehensively improved, reaching a new SOTA perfor- mance without introducing extra computational costs. Complete results can be found in Table 2. dation. Besides, improper exposure will also destroy the image’s structural property and result in detail distortion. The contrast degradation and detail distortion will cause the low-frequency and high-frequency components to mix and inseparable across the image, making the image exposure correction extremely challenging [2, 4, 9, 31, 33, 39]. In practice, one solution for this problem is to design an end-to-end architecture for learning contrast enhancement and detail restoration in shared feature space [14,16]. How- ever, the contrast-relevant features are primarily distributed in low-frequency components, while the detail-relevant fea- tures are primarily distributed in high-frequency compo- nents. Since low-frequency components are statistically dominant over high-frequency components, these methods mainly focus on contrast enhancement and cannot guarantee that the high-frequency details can be efficiently restored. To achieve better contrast enhancement and detail restoration, some researchers propose to decompose and re- store the input image’s lightness and structure components, respectively [2, 20, 41]. For example, some researchers de- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 18115 compose images into illumination and reflectance compo- nents by utilizing Retinex theory and then design a spe- cific network for each component [19, 20, 24]. Other re- searchers propose to decompose the input image into multi- scale components and adopt the coarse-to-fine strategy to progressively recover the lightness and fine-scale structures [2]. However, the decomposition operation inevitably de- stroys the relationship between brightness and detail and cannot balance the contrast and detail enhancement, leading to over-smooth problems or artifacts for enhanced results. To address the above issues, this paper proposes to de- couple the contrast enhancement and detail restoration dur- ing the convolution process. This method is based on statistical observations that the feature response in local regions can be decomposed into low-frequency compo- nents and high-frequency components by a difference oper- ation. Based on this, we introduce a novel Contrast Aware (CA) unit in parallel with a Detail Aware (DA) unit to guide the contrast and detail modeling, termed Decoupling- and-Aggregating Convolution (DAConv). Different from TConv, we inject the addition/difference operation into the convolution process, which can guide the contrast and detail modeling in an explicit manner. Furthermore, to balance the contrast enhancement and detail restoration, we introduce a dynamic coefficient for each branch to adjust the amplitude of the feature response. Our proposed DAConv can be used as a general unit to substitute the TConv kernel in existing CNN-based exposure correction networks to facilitate con- trast enhancement and detail restoration. To reduce the computational costs, the CA, DA, and dy- namic coefficients are aggregated into a single TConv ker- nel by structural re-parameterization in the inference phase. The aggregation is conducted before the activation func- tion, and the linear superposition can reduce computational costs without changing the function of DAConv. After that, the performance of networks can be significantly im- proved without introducing extra computational costs com- pared with the original network. Evaluations of nine meth- ods and five benchmark datasets demonstrate the effective- ness of our proposed method, as shown in Fig. 1. The contribution can be summarized as follows: (1) We propose a novel decoupling-and-aggregating scheme for image exposure correction, in which two par- allel convolution processes are decoupled for contrast en- hancement and detail restoration, respectively, and then ag- gregated into a single branch without additional computa- tion compared with the original convolution scheme. (2) To facilitate the contrast and detail relevant features extraction, a novel CA and DA unit are devised by injecting the addition and difference operation into the convolution process. Compared with traditional convolution kernels, our proposed CA and DA can explicitly model the contrast and detail relevant properties.(3) Evaluations on the five prevailing benchmark datasets and nine SOTA image exposure correction methods demon- strate our proposed DC can comprehensively improve the contrast enhancement and detail restoration performances without introducing extra computational costs.
Wang_Are_We_Ready_for_Vision-Centric_Driving_Streaming_Perception_The_ASAP_CVPR_2023	Abstract In recent years, vision-centric perception has flourished in various autonomous driving tasks, including 3D detec- tion, semantic map construction, motion forecasting, and depth estimation. Nevertheless, the latency of vision-centric approaches is too high for practical deployment ( e.g., most camera-based 3D detectors have a runtime greater than 300ms). To bridge the gap between ideal researches and real-world applications, it is necessary to quantify the trade-off between performance and efficiency. Tradition- ally, autonomous-driving perception benchmarks perform theoffline evaluation, neglecting the inference time de- lay. To mitigate the problem, we propose the Autonomous- driving StreAming Perception (ASAP) benchmark, which is the first benchmark to evaluate the online performance of vision-centric perception in autonomous driving. On the basis of the 2Hz annotated nuScenes dataset, we first propose an annotation-extending pipeline to generate high- frame-rate labels for the 12Hz raw images. Referring to the practical deployment, the Streaming Perception Under const Rained-computation (SPUR) evaluation protocol is further constructed, where the 12Hz inputs are utilized for streaming evaluation under the constraints of differ- ent computational resources. In the ASAP benchmark, comprehensive experiment results reveal that the model rank alters under different constraints, suggesting that the model latency and computation budget should be consid- ered as design choices to optimize the practical deployment. To facilitate further research, we establish baselines for camera-based streaming 3D detection, which consistently enhance the streaming performance across various hard- ware. ASAP project page: https://github.com/ JeffWang987/ASAP .	1. Introduction Vision-centric perception in autonomous driving has drawn extensive attention recently, as it can obtain richer 1Corresponding author, xingang.wang@ia.ac.cn inf TFLOPS @Offline35.6TFLOPS @RTX30909.1TFLOPS @RTX2070S4.4TFLOPS @GTX1060 Computation platform0.10.20.30.4mAP-S FCOS3D PGD BEVDet BEVDet4D BEVFormer PETR BEVDepth BEVDepth-Sv (Ours)Figure 1. Comparison of streaming performances on the ASAP benchmark, where the model rank changes under different com- putational resources. Note that our baseline BEVDepth-Sv (built upon [30]) consistently improves the streaming performance on different platforms. semantic information from images with a desirable budget, compared to LiDAR-based perception. Notably, the past years have witnessed the blooming of vision-centric per- ception in various autonomous driving tasks ( e.g., 3D de- tection [21,22,29–31,33,34,41,56,61], semantic map con- struction [28, 40, 43, 45, 64, 72], motion forecasting [1, 20], and depth estimation [16, 17, 55, 57, 58, 60]). Despite the growing research interest in vision-centric approaches, the high latency of these methods still prevents the practical deployment. Specifically, in the fundamen- tal task of autonomous-driving perception ( e.g., 3D detec- tion), the inference time of most camera-based 3D detec- tors [21,29–31,34,61,70] is longer than 300ms (on the pow- erful NVIDIA RTX3090), which is ∼6×longer (see Tab. 1) than the LiDAR-based counterparts [25, 62, 66]. To enable practical vision-centric perception in autonomous driving, a quantitative metric is in an urgent need to balance the ac- curacy and latency. However, previous autonomous-driving benchmarks [3, 4, 7, 12, 13, 23, 38, 46, 49, 59, 67] mainly fo- This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore. 9600 Table 1. Comparison between autonomous-driving perception dataset, where L&C represents LiDAR and camera, # sensors de- notes number of sensors, Ann. frequency is the annotation fre- quency, and Model speed denotes the typical inference speed of the model on RTX3090. For 2D detectors [2, 11, 15, 44, 50], they achieve ∼45mAP@30FPS on COCO [32]. For LiDAR-based 3D detectors [25, 62, 66], they achieve ∼70mAP@20FPS on Waymo [49]. For camera-based 3D detectors [21, 30, 31, 70], they achieve ∼40mAP@3FPS on nuScenes [4], which is 6 ×∼10×slower than the above two tasks. Dataset Stream. Modality #sensors TaskAnn. Model frequency speed KITTI [12] % L&C - Multi-task - - Argoverse [59] % L&C - Multi-task - - Waymo [49] % L&C - Multi-task - - nuScenes [4] % L&C - Multi-task - - Argoverse-HD [27] ! C 1 2D det. 30Hz ∼30FPS Waymo [18] ! L 1 L-3D det. 10Hz ∼20FPS nuScenes-H ! C 6 C-3D det. 12Hz ∼3FPS cus on the offline performance metrics ( e.g., Average Pre- cision (AP), Truth Positive (TP)), and the model latency has not been particularly studied. Although [18, 27] lever- age the streaming perception paradigm [27] to measure the accuracy-latency trade-off, these benchmarks are designed for 2D detection or LiDAR-based 3D detection. To address the aforementioned problem, this paper proposes the Autonomous-driving StreAming Perception (ASAP) benchmark. To the best of our knowledge, this is the first benchmark to evaluate the online performance of vision-centric perception in autonomous driving. The ASAP benchmark is instantiated on the camera-based 3D detection, which is the core task of vision-centric percep- tion in autonomous driving. To enable the streaming evalua- tion of 3D detectors, an annotation-extending pipeline is de- vised to increase the annotation frame rate of the nuScenes dataset [4] from 2Hz to 12Hz. The extended dataset, termed nuScenes-H (High-frame-rate annotation), is uti- lized to evaluate 3D detectors with 12Hz streaming inputs. Referring to the practical deployment, we delve into the problem of ASAP under different computational resources. Specifically, the Streaming Perception Under const Rained- computation (SPUR) evaluation protocol is constructed: (1) To compare the model performance on varying platforms, multiple GPUs with different computation performances are assigned for the streaming evaluation. (2) To analyze the performance fluctuation caused by the sharing of com- putational resources [9,61,65,70], the streaming evaluation is performed while the GPU is simultaneously processing other perception tasks. As depicted in Fig. 1, the streaming performances of different methods drop steadily as the com- putation power is increasingly constrained. Besides, the model rank alters under various hardware constraints, sug- gesting that the offline performance cannot serve as the de- terministic criterion for different approaches. Therefore, it is necessary to introduce our streaming paradigm to vision-centric driving perception. Based on the ASAP bench- mark, we further establish simple baselines for camera- based streaming 3D detection, and experiment results show that forecasting the future state of the object can compen- sate for the delay in inference time. Notably, the proposed BEVDepth-Sv improves the streaming performance (mAP- S) by∼2%,∼3%, and ∼16% on three GPUs (RTX3090, RTX2070S, GTX1060). The main contributions are summarized as follows: (1) We propose the ASAP benchmark to quantitatively eval- uate the accuracy-latency trade-off of camera-based per- ception methods, which takes a step towards the practical vision-centric perception in autonomous driving. (2) An annotation-extending pipeline is proposed to annotate the 12Hz raw images of the popular nuScenes dataset, which facilitates the streaming evaluation on camera-based 3D de- tection. (3) Simple baselines are established in the ASAP benchmark, which alleviates the influence of inference de- lay and consistently improves the streaming performances across different hardware. (4) The SPUR evaluation proto- col is constructed to facilitate the evaluation of practical de- ployment, where we investigate the streaming performance of the proposed baselines and seven modern camera-based 3D detectors under various computational constraints.