Datasets:

bskang

/

CVPR2023_title_abstract_intro

like 0

Li_SGLoc_Scene_Geometry_Encoding_for_Outdoor_LiDAR_Localization_CVPR_2023

Abstract LiDAR-based absolute pose regression estimates the global pose through a deep network in an end-to-end manner, achieving impressive results in learning-based localization. However, the accuracy of existing methods still has room to improve due to the difficulty of effectively encoding the scene geometry and the unsatisfactory quality of the data. In this work, we propose a novel LiDAR localization frame- work, SGLoc, which decouples the pose estimation to point cloud correspondence regression and pose estimation via this correspondence. This decoupling effectively encodes the scene geometry because the decoupled correspondence regression step greatly preserves the scene geometry, lead- ing to significant performance improvement. Apart from this decoupling, we also design a tri-scale spatial feature aggre- gation module and inter-geometric consistency constraint loss to effectively capture scene geometry. Moreover, we empirically find that the ground truth might be noisy due to GPS/INS measuring errors, greatly reducing the pose estimation performance. Thus, we propose a pose quality evaluation and enhancement method to measure and cor- rect the ground truth pose. Extensive experiments on the Oxford Radar RobotCar and NCLT datasets demonstrate the effectiveness of SGLoc, which outperforms state-of-the-art regression-based localization methods by 68.5% and 67.6% on position accuracy, respectively.

1. Introduction Estimating the position and orientation of LiDAR from point clouds is a fundamental component of many applica- tions in computer vision, e.g., autonomous driving, virtual reality, and augmented reality. Contemporary state-of-the-art LiDAR-based localization methods explicitly use maps, which match the query point *Equal contribution. †Corresponding author. Figure 1. LiDAR Localization results of our method and PosePN++ [51] (state-of-the-art method) in urban (left) and school (right) scenes from Oxford Radar RobotCar [2] and NCLT [34] datasets. The star indicates the starting position. cloud with a pre-built 3D map [18, 23,27,49]. However, these methods usually require expensive 3D map storage and communication. One alternative is the regression- based approach, absolute pose regression (APR), which di- rectly estimates the poses in the inference stage without maps [8, 24,25,40,45]. APR methods typically use a CNN to encode the scene feature and a multi-layer perceptron to regress the pose. Compared to map-based methods, APR does not need to store the pre-built maps, accordingly reduc- ing communications. For (1), APR networks learn highly abstract global scene representations, which allow the network to classify the scene effectively [25]. However, the global features usually cannot encode detailed scene geometry, which is the key to achieving an accurate pose estimation [10, 11,38,39]. Prior efforts have tried to minimize the relative pose or photometric errors to add geometry constraints by pose graph optimization (PGO) [4, 21] or novel view synthesis (NVS) [10, 11]. However, this introduces additional computa- tions, limiting its wide applications. For (2), we empirically find current large-scale outdoor datasets suffer from various errors in the data due to GPS/INS measuring errors. It affects the APR learning process and makes it difficult to evaluate the localization results accurately. To our knowledge, the impact of data quality on localization has not been carefully investigated in the existing literature. 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. 9286 This paper proposes a novel framework, SGLoc, which can (1) effectively capture the scene geometry; In addition, we propose a data pre-processing method, Pose Quality Eval- uation and Enhancement (PQEE), which can (2) improve data quality. (1) Existing APR methods conduct end-to- end regression from the point cloud in LiDAR coordinate to pose. Unlike them, SGLoc decouples this process to (a) regression from the point cloud in LiDAR coordinate to world coordinate and (b) pose estimation via the point cloud correspondence in LiDAR and world coordinate us- ing RANSAC [17]. Importantly, step (a) can effectively preserve the scene geometry, which is key for pose estima- tion [10, 11,38,39]. To achieve high accuracy in step (a), we design a Tri-scale Spatial Feature Aggregation (TSFA) mod- ule and an Inter-Geometric Consistency Constraint (IGCC) loss to effectively capture scene geometry. (2) We empir- ically find that pose errors in the data greatly degrade the pose estimation performance. For example, the ground truth pose obtained by GPS/INS suffers from measuring errors. To address this problem, we proposed a PQEE method which can measure the errors in the pose and correct them after- ward. We conduct extensive experiments on Oxford Radar RobotCar [2] and NCLT [34] datasets, and results show that our method has great advantages over the state-of-the-art, as demonstrated in Fig. 1. Our contributions can be summarized as follows: •SGLoc is the first work to decouple LiDAR localization into point cloud correspondences regression and pose estimation via predicted correspondences, which can ef- fectively capture scene geometry, leading to significant performance improvement. •We propose a novel Tri-Scale Spatial Feature Aggre- gation (TSFA) module and an Inter-Geometric Consis- tency Constraint (IGCC) loss to further improve the encoding of scene geometry. •We propose a generalized pose quality evaluation and enhancement (PQEE) method to measure and correct the pose errors in the localization data, improving 34.2%/16.8% on position and orientation for existing LiDAR localization methods. •Extensive experiments demonstrate the effectiveness of SGLoc, which outperforms state-of-the-art LiDAR localization methods by 68.1% on position accuracy. In addition, to our knowledge, we are the first to reduce the error to the level of the sub-meter on some trajectories.

Li_One-to-Few_Label_Assignment_for_End-to-End_Dense_Detection_CVPR_2023

Abstract One-to-one (o2o) label assignment plays a key role for transformer based end-to-end detection, and it has been re- cently introduced in fully convolutional detectors for end- to-end dense detection. However, o2o can degrade the fea- ture learning efficiency due to the limited number of posi- tive samples. Though extra positive samples are introduced to mitigate this issue in recent DETRs, the computation of self- and cross- attentions in the decoder limits its practi- cal application to dense and fully convolutional detectors. In this work, we propose a simple yet effective one-to-few (o2f) label assignment strategy for end-to-end dense de- tection. Apart from defining one positive and many neg- ative anchors for each object, we define several soft an- chors, which serve as positive and negative samples simul- taneously. The positive and negative weights of these soft anchors are dynamically adjusted during training so that they can contribute more to “representation learning” in the early training stage, and contribute more to “duplicated prediction removal” in the later stage. The detector trained in this way can not only learn a strong feature representa- tion but also perform end-to-end dense detection. Exper- iments on COCO and CrowdHuman datasets demonstrate the effectiveness of the o2f scheme. Code is available at https://github.com/strongwolf/o2f .

1. Introduction Object detection [31, 36, 38, 48] is a fundamental com- puter vision task, aiming to localize and recognize the ob- jects of predefined categories in an image. Owing to the rapid development of deep neural networks (DNN) [14–16, 41,44–46], the detection performance has been significantly improved in the past decade. During the evolution of object detectors, one important trend is to remove the hand-crafted components to achieve end-to-end detection. One hand-crafted component in object detection is the design of training samples. For decades, anchor boxes have *Corresponding author. A B푙￿￿￿=−푡×푙표푔푝−(1−푡)×푙표푔(1−푝) DCA B C D o2o o2f o2m o2o o2f o2mEarly stage LaterstageFigure 1. The positive and negative weights of different anchors (A, B, C and D) in the classification loss during early and later training stages. Each anchor has a positive loss weight t(in orange color) and a negative loss weight 1−t(in blue color). In our method, A is a fully positive anchor, D is a fully negative anchor, and B and C are ambiguous anchors. One can see that for o2o and o2m label assignment schemes, the weights for all anchors are fixed during the training process, while for our o2f scheme, the weights for ambiguous anchors are dynamically adjusted. been dominantly used in modern object detectors such as Faster RCNN [38], SSD [31] and RetinaNet [28]. How- ever, the performance of anchor-based detectors is sensitive to the shape and size of anchor boxes. To mitigate this is- sue, anchor-free [19,48] and query-based [5,8,34,61] detec- tors have been proposed to replace anchor boxes by anchor points and learnable positional queries, respectively. Another hand-crafted component is non-maximum sup- pression (NMS) to remove duplicated predictions. The ne- cessity of NMS comes from the one-to-many (o2m) label assignment [4,13,18,25,26,58], which assigns multiple pos- itive samples to each GT object during the training process. This can result in duplicated predictions in inference and impede the detection performance. Since NMS has hyper- parameters to tune and introduces additional cost, NMS- free end-to-end object detection is highly desired. With a transformer architecture, DETR [5] achieves competitive end-to-end detection performance. Subsequent studies [43, 51] find that the one-to-one (o2o) label as- signment in DETR plays a key role for its success. Con- sequently, the o2o strategy has been introduced in fully convolutional network (FCN) based dense detectors for lightweight end-to-end detection. However, o2o can im- pede the training efficiency due to the limited number 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. 7350 positive samples. This issue becomes severe in dense de- tectors, which usually have more than 10k anchors in an im- age. What’s more, two semantically similar anchors can be adversely defined as positive and negative anchors, respec- tively. Such a ‘label conflicts’ problem further decreases the discrimination of feature representation. As a result, the performance of end-to-end dense detectors still lags behind the ones with NMS. Recent studies [7, 17, 22] on DETR try to overcome this shortcoming of o2o scheme by introducing independent query groups to increase the number of positive samples. The independency between different query groups is ensured by the self-attention computed in the decoder, which is however infeasible for FCN-based detectors. In this paper, we aim to develop an efficient FCN-based dense detector, which is NMS-free yet end-to-end trainable. We observe that it is inappropriate to set the ambiguous an- chors that are semantically similar to the positive sample as fully negative ones in o2o. Instead, they can be used to com- pute both positive and negative losses during training, with- out influencing the end-to-end capacity if the loss weights are carefully designed. Based on the above observation, we propose to assign dynamic soft classification labels for those ambiguous anchors. As shown in Fig. 1, unlike o2o which sets an ambiguous anchor (anchor B or C) as a fully negative sample, we label each ambiguous anchor as par- tially positive and partially negative. The degrees of positive and negative labels are adaptively adjusted during training to keep a good balance between ‘representation learning’ and ‘duplicated prediction removal’. In particular, we be- gin with a large positive degree and a small negative degree in the early training stage so that the network can learn the feature representation ability more efficiently, while in the later training stage, we gradually increase the negative de- grees of ambiguous anchors to supervise the network learn- ing to remove duplicated predictions. We name our method as a one-to-few (o2f) label assignment since one object can have a few soft anchors. We instantiate the o2f LA into dense detector FCOS, and our experiments on COCO [29] and CrowHuman [40] demonstrate that it achieves on-par or even better performance than the detectors with NMS.

Luo_Leverage_Interactive_Affinity_for_Affordance_Learning_CVPR_2023

Abstract Perceiving potential “action possibilities” (i.e., affor- dance) regions of images and learning interactive func- tionalities of objects from human demonstration is a chal- lenging task due to the diversity of human-object inter- actions. Prevailing affordance learning algorithms often adopt the label assignment paradigm and presume that there is a unique relationship between functional region and affordance label, yielding poor performance when adapt- ing to unseen environments with large appearance varia- tions. In this paper, we propose to leverage interactive affin- ity for affordance learning, i.e.extracting interactive affinity from human-object interaction and transferring it to non- interactive objects. Interactive affinity, which represents the contacts between different parts of the human body and lo- cal regions of the target object, can provide inherent cues of interconnectivity between humans and objects, thereby re- ducing the ambiguity of the perceived action possibilities. Specifically, we propose a pose-aided interactive affinity learning framework that exploits human pose to guide the network to learn the interactive affinity from human-object interactions. Particularly, a keypoint heuristic perception (KHP) scheme is devised to exploit the keypoint association of human pose to alleviate the uncertainties due to interac- tion diversities and contact occlusions. Besides, a contact- driven affordance learning (CAL) dataset is constructed by collecting and labeling over 5,000images. Experimental results demonstrate that our method outperforms the rep- resentative models regarding objective metrics and visual quality. Code and dataset: github.com/lhc1224/PIAL-Net.

1. Introduction The objective of affordance learning is to locate the “ac- tion possibilities” regions of an object [15, 18]. For an in- *Corresponding author.‡Equal contributions. Figure 1. (a)Interaction affinity refers to the contact between dif- ferent parts of the human body and the local regions of a target object. (b)The interactive affinity provides rich cues to guide the model to acquire invariant features of the object’s local regions in- teracting with the body part, thus counteracting the multiple pos- sibilities caused by diverse interactions. telligent agent, it is vital to perceive not only the object se- mantics but also how to interact with various objects’ local regions. Perceiving and reasoning about the object’s inter- actable regions is a critical capability for embodied intelli- gent systems to interact with the environment actively, dis- tinct from passive perception systems [3, 38, 39, 44]. More- over, affordance learning has a wide range of applications in fields such as action recognition [13, 24, 43], scene un- derstanding [9, 69], human-robot interaction [51, 63], au- tonomous driving [7] and VR/AR [50, 53]. Affordance is a dynamic property closely related to hu- mans and the environment [18]. Previous works [11,37,40, 46] focus on establishing mapping relationships between appearances and labels for affordance learning. However, they neglect the multiple possibilities of affordance brought about by changes in the environment and actors, leading to an incorrect perception. Recent studies [39, 48] uti- lize reinforcement learning to allow intelligent agents to perceive the environment through numerous interactions in simulated/actual scenarios. Such approaches are mainly limited by their high cost and struggle to generalize 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. 6809 Figure 2. Motivation. (a) This paper explores the associations of interactable regions between diverse images by considering the context of contact regions with different body parts. (b)This paper considers leveraging the connection of human pose keypoints to alleviate the uncertainties due to interaction diversities and contact occlusions. unseen scenarios [58]. To this end, researchers consider learning from human demonstration in an action-free man- ner [14, 29, 30, 38]. Nonetheless, they only roughly seg- ment the whole object/interaction regions in a general way, which is still challenging to understand how the object is used. The multiple possibilities due to different local re- gions interacting with humans in various ways are not fully resolved. In this paper, we propose to leverage interac- tive affinity for affordance learning, i.e.extracting interac- tive affinity from human-object interaction and transferring it to non-interactive objects. The interactive affinity (Fig. 1 (a)) denotes the contacts between different human body parts and objects’ local regions, which can provide inher- ent cues of interconnectivity between humans and objects, thereby reducing the ambiguity of the perceived action pos- sibilities (Fig. 1 (b)). However, it faces the challenges of interaction diversi- ties and contact occlusions, leading to difficulties in acquir- ing a good interactive affinity representation. The human pose is independent of background, and the same interac- tion corresponds to approximately similar poses. Thus, it makes sense to use the association between pose keypoints to overcome the difficulty of obtaining interactive affinity representations. Moreover, it is challenging to transfer the interactive affinity to non-interactive object images due to variations in views, scales, and appearances. The context between the different body part contact regions (Fig. 2 (a)) provides the model with the possibility to explore the as- sociations between the interactable regions of the various images to counteract transfer difficulties. In this paper, we present a pose-aided interactive affin- ity learning framework. First, an Interactive Feature Enhancement ( IFE) module is introduced to explore the connections between different interactable regions of the images. Then, a Keypoint Heuristic Perception ( KHP ) scheme is devised to mine the interactive affinity repre- sentation from interaction and transfer it to non-interactiveobjects. Specifically, the IFE module leverages the trans- former to fully extract global contextual cues by exploiting the common relationships between their local interactable regions (Fig. 2 (a)). Then, they are used to establish asso- ciations between the object interactable regions in different images. Subsequently, the KHP scheme exploits the corre- lation between the human body keypoints and the contact region to guide the network to mine the object’s local in- variant features interacting with the body parts (Fig. 2 (b)). Although the numerous related datasets [9, 10, 19, 30, 36, 57, 67] that emerged during the development of affor- dance learning, there is still a lack of relevant datasets suited for leveraging interactive affinity. To carry out a thorough study, this paper constructs an Contact-driven Affordance Learning ( CAL ) dataset, consisting of 5,258images from 23affordance and 47object categories. We conduct con- trastive studies on the CAL dataset against six representa- tive models in several related fields. Experimental results validate the effectiveness of our method in solving the mul- tiple possibilities of affordance. Contributions: 1) We propose leveraging interactive affinity for affordance learning and establishing a CAL benchmark to facilitate the study of obtaining interactive affinity to counteract the multiple possibilities of affor- dance. 2)We propose a pose-aided interactive affinity learn- ing framework that exploits pose data to guide the network to mine the interactive affinity of body parts and object re- gions from human-object interaction. 3)Experiments on the CAL dataset demonstrate that our model outperforms state-of-the-art methods and can serve as a strong baseline for future affordance learning research.

Nauta_PIP-Net_Patch-Based_Intuitive_Prototypes_for_Interpretable_Image_Classification_CVPR_2023

Abstract Interpretable methods based on prototypical patches rec- ognize various components in an image in order to explain their reasoning to humans. However, existing prototype- based methods can learn prototypes that are not in line with human visual perception, i.e., the same prototype can refer to different concepts in the real world, making interpretation not intuitive. Driven by the principle of explainability-by- design, we introduce PIP-Net (Patch-based Intuitive Proto- types Network): an interpretable image classification model that learns prototypical parts in a self-supervised fashion which correlate better with human vision. PIP-Net can be interpreted as a sparse scoring sheet where the pres- ence of a prototypical part in an image adds evidence for a class. The model can also abstain from a decision for out-of- distribution data by saying “I haven’t seen this before”. We only use image-level labels and do not rely on any part an- notations. PIP-Net is globally interpretable since the set of learned prototypes shows the entire reasoning of the model. A smaller local explanation locates the relevant prototypes in one image. We show that our prototypes correlate with ground-truth object parts, indicating that PIP-Net closes the “semantic gap” between latent space and pixel space. Hence, our PIP-Net with interpretable prototypes enables users to interpret the decision making process in an intuitive, faithful and semantically meaningful way. Code is available athttps://github.com/M-Nauta/PIPNet .

1. Introduction Deep neural networks are dominant in computer vision, but there is a high demand for understanding the reasoning of such complex models [23,30]. Consequently, interpretability and explainability have grown in importance. In contrast to the common post-hoc explainability that reverse-engineersa black box, we argue that we should take interpretability as a design starting point for in-model explainability. The recognition-by-components theory [1] describes how hu- mans recognize objects by segmenting them into multiple components. We mimic this intuitive line of reasoning in an intrinsically interpretable image classifier. Specifically, our PIP-Net (Patch-based Intuitive Prototypes Network) au- tomatically identifies semantically meaningful components, while only having access to image-level class labels and not relying on additional part annotations. The components are “prototypical parts” (prototypes) visualized as image patches, since exemplary natural images are more informa- tive to humans than generated synthetic images [2]. PIP-Net is globally interpretable and designed to be highly intuitive as it uses simple scoring-sheet reasoning: the more relevant prototypical parts for a specific class are present in an image, the more evidence for that class is found, and the higher its score. When no relevant prototypes are present in the image, with e.g. out-of-distribution data, PIP-Net will abstain from a decision. PIP-Net is therefore able to say “I haven’t seen this before” (see Fig. 2). Additionally, following the principle of isolation of functional properties for aligning human and machine vision [5], the reasoning of PIP-Net is separated into multiple steps. This simplifies human identification of reasons for (mis)classification. Recent interpretable part-prototype models are ProtoP- Net [3], ProtoTree [24], ProtoPShare [29] and ProtoPool [28]. These part-prototype models are only designed for fine- grained image recognition tasks (birds and car types) and lack “semantic correspondence” [17] between learned proto- types and human concepts. This “semantic gap” in prototype- based methods between similarity in latent space and input space was also found by others [9, 14]. We hypothesize that the main cause of the semantic gap is the fact that exist- ing part-prototype models only regularize interpretability on class-level, since their underlying assumption is that (parts 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. 2744 Sun OR DogNOT (sun OR dog) Image patches similar to prototype 1 Image patches similar to prototype 1 (“dog”)Existing prototype-based modelsOur model Image patches similar to prototype 2 (“sun”) Figure 1. Toy dataset with two classes (left). Existing models can learn representations of prototypes that do not align with human visually perceived similarity (center). Our objective is to learn prototypes that represent concepts that also look similar to humans (right). of) images from the same class have the same prototypes. This assumption may however not hold, leading to similarity in latent space which does not correspond to visually per- ceived similarity. Consider the example in Fig. 1, where we have re-labeled images from a clipart dataset [43] to create a binary classification task: the two kids are happy when the sun or dog is present, and sad when there is neither a sun nor a dog. Hence, the classes are ‘ sun OR dog ’ and ‘ NOT (sun OR dog) ’. Intuitively, an easy-to-interpret model should learn two prototypes: one for the sun and one for the dog. However, existing interpretable part-prototype models, such as ProtoPNet [3] and ProtoTree [24], optimize images of the same class to have the same prototypes. They could, there- fore, learn a single prototype that represents both the sun and the dog, especially when the model is optimized to have few prototypes (see Fig. 1, center). The model’s perception of patch similarity may thus not be in line with human visual perception, leading to the perceived “semantic gap”. To address the gap between latent and pixel space, we present PIP-Net: an interpretable model that is designed to be intuitive and optimized to correlate with human vision. A sparse linear layer connects learned interpretable proto- typical parts to classes. A user only needs to inspect the prototypes and their relation to the classes in order to inter- pret the model. We restrict the weights of the linear layer to be non-negative, such that the presence of a class-relevant prototype increases the evidence for a class. The linear layer can be interpreted as a scoring sheet: the score for a class is the sum of all present prototypes multiplied by their weights. Alocal explanation (Fig. 2 and Fig. 3) explains a specific prediction and shows which prototypes were found at which locations in the image. The global explanation provides an overall view of the model’s decision layer, consisting of the sparse weights between classes and their relevant prototypes. Because of this interpretable andpredictive linear layer, we ensure a direct relation between the prototypes and the clas- sification, and thereby prevent unfaithful explanations which can arise with local or post-hoc XAI methods [16].Our Contributions: 1.We present the Patch-based Intuitive Prototypes Net- work (PIP-Net): an intrinsically interpretable image classifier, driven by three explainability requirements: the model should be intuitive, compact and able to han- dle out-of-distribution data. 2.PIP-Net has a surprisingly simple architecture and is trained with novel regularization for learning prototype similarity that better correlates with human visual per- ception , thereby closing a perceived semantic gap. 3.PIP-Net acts as a scoring sheet and therefore can detect that an image does not belong to anyclass or that it belongs to multiple classes. 4.Instead of specifying the number of prototypes be- forehand as in ProtoPNet [3], ProtoPool [28] and Tes- Net [35], PIP-Net only needs an upper bound on the number of prototypes and selects as few prototypes as possible for good classification accuracy with compact explanations, reaching sparsity ratios >99%.

Ma_DiGeo_Discriminative_Geometry-Aware_Learning_for_Generalized_Few-Shot_Object_Detection_CVPR_2023

Abstract Generalized few-shot object detection aims to achieve precise detection on both base classes with abundant an- notations and novel classes with limited training data. Ex- isting approaches enhance few-shot generalization with the sacrifice of base-class performance, or maintain high pre- cision in base-class detection with limited improvement in novel-class adaptation. In this paper, we point out the rea- son is insufficient Di scriminative feature learning for all of the classes. As such, we propose a new training frame- work, DiGeo, to learn Geo metry-aware features of inter- class separation and intra-class compactness. To guide the separation of feature clusters, we derive an offline sim- plex equiangular tight frame (ETF) classifier whose weights serve as class centers and are maximally and equally sep- arated. To tighten the cluster for each class, we include adaptive class-specific margins into the classification loss and encourage the features close to the class centers. Ex- perimental studies on two few-shot benchmark datasets (VOC, COCO) and one long-tail dataset (LVIS) demon- strate that, with a single model, our method can effectively improve generalization on novel classes without hurting the detection of base classes. Our code can be found here.

1. Introduction Recent years have witnessed the tremendous growth of object detection through deep neural models and large-scale training [2, 13–15, 40, 42, 45, 60, 65]. However, the success of detection models heavily relies on the amount and qual- ity of annotations, which requires expensive annotation cost and time. In addition, traditional object detection models perform worse on the classes with a limited number of an- notations [11, 52, 56], while human are able to learn from few observations. In order to close the gap between human vision system and detection models, recent studies have investigated how to generalize well on rare classes under Corresponding Author. 10-shot5-shot3-shotNovel class detection (nAP50) 4055708530405060 20Base class detection (bAP50)FsDetViewMeta R-CNNFSRWFRCNft-fullTFARetentiveRCNNOursMPSRTransfer LearningOursMeta Learning10-shot5-shot3-shot (base-onlyPre-training)Figure 1. Performance on few-shot object detection on Pascal VOC [3]. Previous transfer-learning approaches (blue) balancing the training data by aggressively down-sampling the base set and may result in overfitting. Instead, we (red) use the full train set, aiming to both maintain precise base detection but learn discrimi- native features from the limited annotations for few-shot classes. the few-shot object detection (FSOD) setting. Specifically, given many-shot ( base) classes with plenty of training data and few-shot ( novel ) classes with extremely limited training data ( e.g., 5 annotated instances per class), FSOD expects the model to detect the objects in the novel classes well. To improve the generalization ability on novel-class de- tection, recent studies [6, 44, 52] conduct transfer learning in a two-step manner. In detail, the model is pre-trained on the whole set of base classes, and then fine-tuned on the union of the set of novel classes and an aggressively down- sampled base subset. However, the efficient few-shot adap- tation is often achieved at the expense of sacrificing preci- sion on base detection (Fig. 1). Being aware of this limita- tion, Fan et al. [6] proposed to evaluate the performance of both base and novel classes in the generalized few-shot ob- ject detection (GFSOD) setting. In addition, they proposed a consistency regularization to emphasize the pre-trained base knowledge during fine-tuning and employed an ensem- bling strategy. However, they design different classifiers for base and novel classes, and the adaptation on novel classes is impeded due to a complex ensembling process. In this paper, we pointed out that the devil is in 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. 3208 sufficient discriminative feature learning for few-shot ob- ject detection, including inefficient knowledge adaptation to novel classes and unexpected knowledge forgetting of base classes. First, as the novel instances are extremely limited during training, it is hard to capture the representative vi- sual information of novel classes and adapt the knowledge learned from base classes to novel classes. As a result, the model cannot distinguish between the novel classes, which weakens the few-shot adaptation. Secondly, balanced train- ing strategies such as down-sampling fail to utilize the di- verse training samples from base set. Thus, it is hard to pre- serve the complete knowledge of base classes, which leads to overfitting and further decreases the detection scores. To tackle these challenges, we proposed a new training framework, DiGeo , to make the best of both worlds for generalized few-shot object detection, i.e., improving gen- eralization on novel classes without hurting the detection of base classes. Our motivation is to learn Discriminative Geometry-aware features via inter-class separation and intra-class compactness . For inter-class separation, we ex- pect the class centers [53] to be well distinct from each other. Motivated by the symmetric geometry of simplex equiangular tight frame (ETF) [36], we proposed to use ETF as classifier to guide the separation of features. To be spe- cific, we derive an offline ETF whose weights are maxi- mally & equivalently separated ( i.e., independent from the training data distribution) and are assigned as fixed centers for all classes. For intra-class compactness, we expect the features to be closed to the class centers for a clear deci- sion boundary. In practice, we add class-specific margins to output logits during training to push the features close to the class centers. The margins are based on instance dis- tribution prior and are then adaptively adjusted though self- distillation. Meanwhile, we consider the huge imbalance between base set and novel set, and up-sample the novel set to facilitate the feature extraction. We validate the effectiveness of DiGeo under the GF- SOD setting on Pascal VOC [3, 4] and MS COCO [27]. Compared to existing methods, we can both achieve pre- cise detection on base classes and sufficiently improve the adaptation efficiency on novel classes using a single model. Furthermore, our DiGeo can be intuitively extended to long-tailed object detection. Experimental results on LVIS datasets demonstrate the generalizibility of our approach. Our contributions are summarized as follows: We revisit few-shot object detection from a perspective of discriminative feature learning, and point out that exist- ing methods fail in knowledge adaptation to novel classes and suffer from knowledge forgetting of base classes. We propose DiGeo to pursue an desired feature geom- etry, i.e., inter-class separation and intra-class compact- ness, which consistently improves the performance on both base and novel classes.We conduct extensive experiments on three benchmark datasets for few-shot object detection and long-tailed ob- ject detection to verify the generalizability of DiGeo.

Qiao_Fuzzy_Positive_Learning_for_Semi-Supervised_Semantic_Segmentation_CVPR_2023

Abstract Semi-supervised learning (SSL) essentially pursues class boundary exploration with less dependence on human an- notations. Although typical attempts focus on ameliorat- ing the inevitable error-prone pseudo-labeling, we think differently and resort to exhausting informative semantics from multiple probably correct candidate labels. In this pa- per, we introduce Fuzzy Positive Learning (FPL) for accu- rate SSL semantic segmentation in a plug-and-play fash- ion, targeting adaptively encouraging fuzzy positive pre- dictions and suppressing highly-probable negatives. Be- ing conceptually simple yet practically effective, FPL can remarkably alleviate interference from wrong pseudo la- bels and progressively achieve clear pixel-level semantic discrimination. Concretely, our FPL approach consists of two main components, including fuzzy positive assign- ment (FPA) to provide an adaptive number of labels for each pixel and fuzzy positive regularization (FPR) to re- strict the predictions of fuzzy positive categories to be larger than the rest under different perturbations. Theo- retical analysis and extensive experiments on Cityscapes and VOC 2012 with consistent performance gain justify the superiority of our approach. Codes are provided in https://github.com/qpc1611094/FPL .

1. Introduction Semantic segmentation models enable accurate scene understanding [1, 29,45] with the help of fine pixel-level annotations. Yet, collecting labeled segmentation datasets is time-consuming and labor-costing [6]. Considering unla- beled data are annotation-free and easily accessible, semi- *Equal contribution. †Corresponding author.supervised learning (SSL) is introduced into semantic seg- mentation [5, 34, 43, 49, 51, 53] to encourage the model to generalize better on unseen data with less dependence on artificial annotations. Figure 1. (a) Existing methods using pseudo label to utilize un- labeled data. (b) The proposed FPL that provides multiple fuzzy positive labels for each pixel to utilize unlabeled data. The exam- ple of ‘Truck’ shows that our method covers ground truth (GT) more comprehensively than vanilla positive learning. The semi-supervised segmentation task faces a scenario where only a subset of training images are assigned seg- mentation labels while the others remain unlabeled. Cur- rent state-of-the-art (SOTA) methods utilize unlabeled data via consistency regularization, which aims to obtain invari- ant predictions for unlabeled pixels under various perturba- tions [5, 34, 49, 53]. Their general paradigm is to use the pseudo label generated under weak (or none) perturbations as the learning target of predictions under strong perturba- tions. Though achieving promising results, errors are in- evitable in the pseudo label used in these methods, misguid- 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. 15465 ing the training of their models [24,33]. An intuitive exam- ple is that some pixels may be confused in categories with similar semantics. As Fig. 1 (a), some pixels belonging to ‘Truck’ are wrongly classified into the ‘Car’ category (e.g., white boxed pixel). To mitigate this problem, typical meth- ods focus on ameliorating the learning of pseudo labels by filtering low-confidence pseudo labels out [14,21,38,51,53] and generating pseudo labels more accurately [8,20,26,48]. However, the semantics of ground truth buried in other un- selected labels are ignored in existing methods. In this paper, we propose Fuzzy Positive Learning (FPL), a new SSL segmentation method that exhausts in- formative semantics from multiple probably correct candi- date labels. We name these labels “fuzzy positive” labels since each of them has the probability to be the ground truth. As shown in Fig. 1 (b), our fuzzy positive labels cover the ground truth more comprehensively, facilitating our FPL to exploit the semantics of ground truth better. Ex- tending learning from one pseudo label to learning from multiple fuzzy positive labels is not a simple implemen- tation, which contains two pending issues. One is how to provide an adaptive number of labels for each pixel. And the other one is how to exploit the possible GT semantics from fuzzy positive labels. For these two issues, a fuzzy positive assignment (FPA) algorithm is first proposed to se- lect which labels should be appended to the fuzzy positive label set of each pixel. Afterward, a fuzzy positive regular- ization (FPR) is developed to regularize the predictions of fuzzy positive categories to be larger than the predictions of the rest negative categories under different perturbations. Our FPL achieves consistent performance gain on Cityscapes and Pascal VOC 2012 datasets using CPS [5] and AEL [14] as baselines. Moreover, we theoretically and empirically analyze that the superiority of FPL lies in re- vising the gradient of learning ground truth when pseudo- labels are wrongly-assigned. Our main contributions are: • FPL provides a new perspective for SSL segmentation, that is, learning informative semantics from multiple fuzzy positive labels instead of only one pseudo label. • A fuzzy positive assignment is proposed to provide an adaptive number of labels for each pixel. Besides, a fuzzy positive regularization is developed to learn the semantics of ground truth from fuzzy positive labels. • FPL is easy to implement and could bring stable per- formance gains on existing SSL segmentation methods in a plug-and-play fashion.

Li_Spatially_Adaptive_Self-Supervised_Learning_for_Real-World_Image_Denoising_CVPR_2023

Abstract Significant progress has been made in self-supervised image denoising (SSID) in the recent few years. However ,most methods focus on dealing with spatially independentnoise, and they have little practicality on real-world sRGB images with spatially correlated noise. Although pixel-shuffle downsampling has been suggested for breaking the noise correlation, it breaks the original information of im-ages, which limits the denoising performance. In this paper , we propose a novel perspective to solve this problem, i.e., seeking for spatially adaptive supervision for real-world sRGB image denoising. Specifically, we take into accountthe respective characteristics of flat and textured regions in noisy images, and construct supervisions for them sepa-rately. F or flat areas, the supervision can be safely derived from non-adjacent pixels, which are much far from the cur-rent pixel for excluding the influence of the noise-correlatedones. And we extend the blind-spot network to a blind-neighborhood network (BNN) for providing supervision onflat areas. F or textured regions, the supervision has to beclosely related to the content of adjacent pixels. And wepresent a locally aware network (LAN) to meet the require-ment, while LAN itself is selectively supervised with the out-put of BNN. Combining these two supervisions, a denoising network ( e.g., U-Net) can be well-trained. Extensive exper- iments show that our method performs favorably against state-of-the-art SSID methods on real-world sRGB pho- tographs. The code is available at https://github. com/nagejacob/SpatiallyAdaptiveSSID .

1. Introduction Image denoising aims to restore clean images from noisy observations [ 5,11,14], and it has achieved noticeable im- provement with the advances in deep networks [ 2,10,21, 29–31,33,38,40,41,46–48,50,51]. However, the mod- els trained with synthetic noise usually perform poorly in real-world scenarios in which noise is complex and change- (a) Noisy Input (b) CVF-SID [ 35] (c) AP-BSN+R3[27] (d) Ours Figure 1. Visual comparison between self-supervised denoising methods on the DND dataset [ 37]. PSNR (dB) and SSIM with respect to the ground-truth are marked on the result for quantitativecomparison. Our method performs better in removing spatiallycorrelated noise from real-world sRGB photographs. able. A feasible solution is to collect real-world clean- noisy image pairs [ 1,37] and take them for model train- ing [ 2,15,21,46,47]. But building such datasets generally requires strictly controlled environment as well as compli- cated photographing and post-processing, which is time-consuming and labor-intensive. Moreover, the noise statis- tics vary under different cameras and illuminating condi- tions [ 43,54], and it is impractical to capture pairs for every device and scenario. To circumvent the limitations of noisy-clean pairs collection, self-supervised image denoising (SSID) 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. 9914 proaches [ 3,17,19,20,24,26–28,34–36,42,45] have been proposed, which can be trained merely on noisy im-ages. However, the noise model assumptions of a largeamount of SSID methods do not match the characteristicsof real-world noise in sRGB space. For instance, HQ- SSL [ 26] improves the denoising performance with poste- rior inference, but requires explicit noise probability den-sity. Noise2Score [ 20] and its extension [ 19] propose a closed-form image denoising schema with score matchingfollowed by noise model and noise level estimation, butthe noise is bounded to Tweedie distribution. Althoughsome methods [ 17,42] are designed for distribution agnostic noise, they can only deal with spatially independent noise. Recently, a few attempts have been explored to remove spatially correlated noise in a self-supervised manner. CVF- SID [ 35] disentangles the image and noise components from noisy images, but the difficulty of optimization lim- its its performance. Some methods [ 44,57] break the spa- tial noise correlation with pixel-shuffle downsampling (PD), then utilize spatially independent denoisers ( e.g., blind-spot network [ 6,26,44]) to remove the uncorrelated noise. How- ever, PD breaks the original information of the images andleads to aliasing artifacts, which largely degrade the imagequality. AP-BSN [ 27] applies asymmetric PD factors and post-refinement processing to seek for a better trade-off be-tween noise removal and aliasing artifacts, but it is time- consuming during inference. In this paper, we present a novel perspective for SSID by considering the respective characteristics of flat and tex-tured regions in noisy images, resulting in a spatially adap- tive SSID method for real-world sRGB images. Insteadof utilizing pixel-shuffle downsampling and blind-spot net-work to learn denoising results directly, we seek for spa- tially adaptive supervision for a denoising network ( e.g., U- Net [ 39]). Concretely, for flat areas, the supervision can be safely derived from non-adjacent pixels, which are much far from the current pixel for excluding the influence of noisecorrelation. We achieve it by extending the blind-spot net- work (BSN) [ 26] to a blind-neighborhood network (BNN). BNN modifies the architecture of BSN to expand the sizeof blind region, and takes the same self-supervised train-ing schema as BSN. Note that it is difficult to determine whether an area is flat or not from the noisy images, so we directly apply BNN to the whole image and it has little ef- fect on the handling of flat areas. Moreover, such an op- eration can give us a chance to detect textured areas fromthe output of BNN, whose variance is usually higher. For textured areas, neighboring pixels are essential for predict-ing the details and they can not be ignored. To this end, we present a locally aware network (LAN), which focuses on recovering the texture details solely from adjacent pix- els. LAN is supervised by flat areas of BNN output. Whentraining is done, LAN will be applied to textured areas togenerate supervision information for these areas. Combining the learned supervisions for flat and textured areas, a denoising network can be readily trained. Dur-ing inference, BNN and LAN can be detached, only theultimate denoising network is used to restore clean im-ages. Extensive experiments are conducted on SIDD [ 1] and DND [ 37] datasets. The results demonstrate our method is not only effective but also efficient. In comparison to state-of-the-art self-supervised denoising methods, our method behaves favorably in terms of both quantitative metrics and perceptual quality. The contributions of this paper can be summarized as follows: • We propose a novel perspective for self-supervised real-world image denoising, i.e., learning spatially adaptive supervision for a denoising network accord- ing to the image characteristics. • For flat areas, we extend the blind-spot network to a blind-neighborhood network (BNN) for providing su- pervision information. For texture areas, we present alocally aware network (LAN) to learn that from neigh- boring pixels. • Extensive experiments show our method has superior performance and inference efficiency against state-of-the-art SSID methods on real-world sRGB noise re- moval.

Muller_DiffRF_Rendering-Guided_3D_Radiance_Field_Diffusion_CVPR_2023

Abstract We introduce DiffRF , a novel approach for 3D radiance field synthesis based on denoising diffusion probabilistic models. While existing diffusion-based methods operate on images, latent codes, or point cloud data, we are the first to directly generate volumetric radiance fields. To this end, we propose a 3D denoising model which directly operates on an explicit voxel grid representation. However, as radi- ance fields generated from a set of posed images can be am- biguous and contain artifacts, obtaining ground truth radi- ance field samples is non-trivial. We address this challenge by pairing the denoising formulation with a rendering loss, enabling our model to learn a deviated prior that favours good image quality instead of trying to replicate fitting er-rors like floating artifacts. In contrast to 2D-diffusion mod- els, our model learns multi-view consistent priors, enabling free-view synthesis and accurate shape generation. Com- pared to 3D GANs, our diffusion-based approach naturally enables conditional generation such as masked completion or single-view 3D synthesis at inference time.

1. Introduction In recent years, Neural Radiance Fields (NeRFs) [37] have emerged as a powerful representation for fitting indi- vidual 3D scenes from posed 2D input images. The ability to photo-realistically synthesize novel views from arbitrary viewpoints while respecting the underlying 3D scene geom- Project page: https://sirwyver.github.io/DiffRF/. 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. 4328 etry has the potential to disrupt and transform applications like AR/VR, gaming, mapping, navigation, etc. A num- ber of recent works have introduced extensions for making NeRFs more sophisticated, by e.g., showing how to incor- porate scene semantics [18, 30], training models from het- erogeneous data sources [35], or scaling them up to repre- sent large-scale scenes [62, 64]. These advances are testa- ment to the versatility of ML-based scene representations; however, they still fit to specific, individual scenes rather than generalizing beyond their input training data. In contrast, neural field representations that general- ize to multiple object categories or learn priors for scenes across datasets appear much more limited to date, despite enabling applications like single-image 3D object gener- ation [7, 39, 49, 66, 72] and unconstrained scene explo- ration [15]. These methods explore ways to disentangle ob- ject priors into shape and appearance-based components, or to decompose radiance fields into several small and locally- conditioned radiance fields to improve scene generation quality; however, their results still leave significant gaps w.r.t. photorealism and geometric accuracy. Directions involving generative adversarial networks (GANs) that have been extended from the 2D domain to 3D- aware neural fields generation are demonstrating impressive synthesis results [8]. Like regular 2D GANs, the training objective is based on discriminating 2D images, which are obtained by rendering synthesized 3D radiance fields. At the same time, diffusion-based models [52] have re- cently taken the computer vision research community by storm, performing on-par or even surpassing GANs on multiple 2D benchmarks, and are producing photo-realistic images that are almost indistinguishable from real pho- tographs. For multi-modal or conditional settings such as text-to-image synthesis, we currently observe unprece- dented output quality and diversity from diffusion-based ap- proaches. While several works address purely geometric representations [33, 75], lifting the denoising-diffusion for- mulation directly to 3D volumetric radiance fields remains challenging. The main reason lies in the nature of diffu- sion models, which require a one-to-one mapping between the noise vector and the corresponding ground truth data samples. In the context of radiance fields, such volumetric ground truth data is practically infeasible to obtain, since even running a costly per-sample NeRF optimization results in incomplete and imperfect radiance field reconstructions. In this work, we present the first diffusion-based gen- erative model that directly synthesizes 3D radiance fields, thus unlocking high-quality 3D asset generation for both shape and appearance. Our goal is to learn such a gener- ative model trained across objects, where each sample is given by a set of posed RGB images. To this end, we propose a 3D denoising model directly operating on an explicit voxel grid representation (Fig. 1,left) producing high-frequency noise estimates. To address the ambiguous and imperfect radiance field representation for each training sample, we propose to bias the noise pre- diction formulation from Denoising Diffusion Probabilistic Models (DDPMs) towards synthesizing higher image qual- ity by an additional volumetric rendering loss on the esti- mates. This enables our method to learn radiance field pri- ors less prone to fitting artifacts or noise accumulation dur- ing the sampling process. We show that our formulation leads to diverse and geometrically-accurate radiance field synthesis producing efficient, realistic, and view-consistent renderings. Our learned diffusion prior can be applied in an unconditional setting where 3D object synthesis is obtained in a multi-view consistent way, generating highly-accurate 3D shapes and allowing for free-view synthesis. We further introduce the new task of conditional masked completion – analog to shape completion – for radiance field completion at inference time. In this setting, we allow for realistic 3D completion of partially-masked objects without the need for task-specific model adaptation or training (see Fig. 1, right). We summarize our contributions as follows: • To the best of our knowledge, we introduce the first diffusion model to operate directly on 3D radiance fields, enabling high-quality, truthful 3D geometry and image synthesis. • We introduce the novel application of 3D radiance field masked completion, which can be interpreted as a nat- ural extension of image inpainting to the volumetric domain. • We show compelling results in unconditional and con- ditional settings, e.g., by improving over GAN-based approaches on image quality (from 16.54 to 15.95 in FID) and geometry synthesis (improving MMD from 5.62 to 4.42), on the challenging PhotoShape Chairs dataset [44].

Peng_Representing_Volumetric_Videos_As_Dynamic_MLP_Maps_CVPR_2023

Abstract This paper introduces a novel representation of volumet- ric videos for real-time view synthesis of dynamic scenes. Recent advances in neural scene representations demon- strate their remarkable capability to model and render com- plex static scenes, but extending them to represent dynamic scenes is not straightforward due to their slow rendering speed or high storage cost. To solve this problem, our key idea is to represent the radiance field of each frame as a set of shallow MLP networks whose parameters are stored in 2D grids, called MLP maps, and dynamically predicted by a 2D CNN decoder shared by all frames. Represent- ing 3D scenes with shallow MLPs significantly improves the rendering speed, while dynamically predicting MLP pa- rameters with a shared 2D CNN instead of explicitly stor- ing them leads to low storage cost. Experiments show that the proposed approach achieves state-of-the-art rendering quality on the NHR and ZJU-MoCap datasets, while being efficient for real-time rendering with a speed of 41.7 fps for 512×512images on an RTX 3090 GPU. The code is avail- able at https://zju3dv.github.io/mlp maps/.

1. Introduction V olumetric video captures a dynamic scene in 3D which allows users to watch from arbitrary viewpoints with im- mersive experience. It is a cornerstone for the next gener- ation media and has many important applications such as video conferencing, sport broadcasting, and remote learn- ing. The same as 2D video, volumetric video should be ca- pable of high-quality and real-time rendering as well as be- ing compressed for efficient storage and transmission. De- signing a proper representation for volumetric video to sat- isfy these requirements remains an open problem. Traditional image-based rendering methods [1,12,25,74] build free-viewpoint video systems based on dense camera arrays. They record dynamic scenes with many cameras and then synthesize novel views by interpolation from in- put nearby views. For these methods, the underlying scene ∗Equal contribution.†Corresponding author. 2D CNN decoderMLP parameters MLP mapslatent vector +Figure 1. The basic idea of dynamic MLP maps. Instead of modeling the volumetric video with a big MLP network [26], we exploit a 2D convolutional neural network to dynamically gener- ate 2D MLP maps at each video frame, where each pixel storing the parameter vector of a small MLP network. This enables us to represent volumetric videos with a set of small MLP networks, thus significantly improving the rendering speed. representation is the original multi-view video. While there have been many multi-view video coding techniques, the storage and transmission cost is still huge which cannot satisfy real-time video applications. Another line of work [11, 13] utilizes RGB-D sensors to reconstruct textured meshes as the scene representation. With mesh compression techniques, this representation can be very compact and en- able streamable volumetric videos, but these methods can only capture humans and objects in constrained environ- ments as reconstructing a high-quality renderable mesh for general dynamic scenes is still a very challenging problem. Recent advances in neural scene representations [26, 33, 61] provide a promising solution for this problem. They represent 3D scene with neural networks, which can be ef- fectively learned from multi-view images through differen- tiable renderers. For instance, Neural V olumes [33] rep- resents volumetric videos with a set of RGB-density vol- umes predicted by 3D CNNs. Since the volume prediction easily consumes large amount of GPU memory, it strug- gles to model high-resolution 3D scenes. NeRF [37] in- stead represent 3D scenes with MLP networks regressing density and color for any 3D point, thereby enabling it to synthesize high-resolution images. DyNeRF [26] extends NeRF to model volumetric videos by introducing a tempo- ral latent code as additional input of the MLP network. A major issue of NeRF models is that their rendering is gen- 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. 4252 erally quite slow due to the costly network evaluation. To increase the rendering speed, some methods [17,61,75] uti- lize caching techniques to pre-compute a discrete radiance volume. This strategy typically leads to high storage cost, which is acceptable for a static scene, but not scalable to render a volumetric video of dynamic scenes. In this paper, we propose a novel representation of volu- metric video, named dynamic MLP maps, for efficient view synthesis of dynamic scenes. The basic idea is illustrated in Figure 1. Instead of modeling a volumetric video with a sin- gle MLP network, we represent each video frame as a set of small MLP networks whose parameters are predicted by a per-scene trained 2D CNN decoder with a per-frame latent code. Specifically, given a multi-view video, we choose a subset of views and feed them into a CNN encoder to obtain a latent code for each frame. Then, a 2D CNN decoder re- gresses from the latent code to 2D maps, where each pixel in the maps stores a vector of MLP parameters. We call these 2D maps as MLP maps. To model a 3D scene with the MLP maps, we project a query point in 3D space onto the MLP maps and use the corresponding MLP networks to infer its density and color values. Representing 3D scenes with many small MLP networks decreases the cost of network evaluation and increases the rendering speed. This strategy has been proposed in pre- vious works [48, 49], but their networks need to be stored for each static scene, which easily consumes a lot of stor- age to represent a dynamic scene. In contrast to them, we use shared 2D CNN encoder and decoder to predict MLP parameters on the fly for each video frame, thereby effec- tively compressing the storage along the temporal domain. Another advantage of the proposed representation is that MLP maps represent 3D scenes with 2D maps, enabling us to adopt 2D CNNs as the decoder instead of 3D CNNs in Neural V olumes [33]. This strategy leverages the fast infer- ence speed of 2D CNNs and further decreases the memory requirement. We evaluate our approach on the NHR and ZJU-MoCap datasets, which present dynamic scenes with complex mo- tions. Across all datasets, our approach exhibits state-of- the-art performance in terms of rendering quality and speed, while taking up low storage. Experiments demonstrate that our approach is over 100 times faster than DyNeRF [26]. In summary, this work has the following contributions: • A novel representation of volumetric video named dy- namic MLP maps, which achieves compact represen- tation and fast inference. • A new pipeline for real-time rendering of dynamic scenes based on dynamic MLP maps. • State-of-the-art performance in terms of the render- ing quality, speed, and storage on the NHR and ZJU- MoCap datasets.

Miao_FedSeg_Class-Heterogeneous_Federated_Learning_for_Semantic_Segmentation_CVPR_2023

Abstract Federated Learning (FL) is a distributed learning paradigm that collaboratively learns a global model by multiple clients with data privacy-preserving. Although many FL algorithms have been proposed for classification tasks, few works focus on more challenging semantic seg- mentation tasks, especially in the class-heterogeneous FL situation. Compared with classification, the issues from het- erogeneous FL for semantic segmentation are more severe: (1) Due to the non-IID distribution, different clients may contain inconsistent foreground-background classes, result- ing in divergent local updates. (2) Class-heterogeneity for complex dense prediction tasks makes the local optimum of clients farther from the global optimum. In this work, we propose FedSeg, a basic federated learning approach for class-heterogeneous semantic segmentation. We first propose a simple but strong modified cross-entropy loss to correct the local optimization and address the foreground- background inconsistency problem. Based on it, we intro- duce pixel-level contrastive learning to enforce local pixel embeddings belonging to the global semantic space. Ex- tensive experiments on four semantic segmentation bench- marks (Cityscapes, CamVID, PascalVOC and ADE20k) demonstrate the effectiveness of our FedSeg. We hope this work will attract more attention from the FL community to the challenging semantic segmentation federated learning.

1. Introduction Semantic segmentation is the task of assigning a unique semantic label to every pixel in a given image, which is a fundamental research topic in computer vision and has many potential applications, such as autonomous driving, image editing and robotics [30]. Training a semantic seg- mentation model usually needs vast of data with pixel-level annotations, which is extremely hard to acquire. Collabo- rative training on multiple clients is a feasible way to solve †Corresponding author. Client 1Client 2Client 3 Server (a) (b)𝒘𝟏𝒕𝒘𝒘𝟏𝒕#𝟏𝒘𝟐𝒕𝒘𝟑𝒕𝒘𝟐𝒕#𝟏𝒘𝟑𝒕#𝟏Pixel Embedding Space (motorbike)(person)(cat) Class-Heterogeneous Dense PredictionClients’Optimization Divergence??????ClientsServerFigure 1. (a) The foreground-background inconsistency for class- heterogeneous semantic segmentation. (b) Local optimization di- vergence problem for the heterogeneous dense prediction task. the problem. However, collaborative training has the risk of leaking sensitive information. For example, for the au- tonomous driving task, the training images may include pri- vate information such as where the user arrived, where the user lives and what the user’s house looks like. Thus, a privacy-preserving collaborative training method is requi- site for semantic segmentation. Federated Learning (FL) [31] is an emerging distributed machine learning paradigm that jointly trains a shared global model by multiple clients without exchanging their raw data. FedAvg [31] is a basic FL algorithm that learns local models with raw data on clients separately while ag- gregating weights to a global model on a server. One key problem of FL is the statistical heterogeneity of data dis- tribution among different clients. Many recent FL algo- rithms [1, 21, 22, 26, 32] are proposed to tackle the prob- lem. However, most of them evaluate their methods on classification, while few works focus on more challeng- ing semantic segmentation. Although some federated learn- 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. 8042 ing approaches [17, 29, 46, 52] for medical image segmen- tation have been proposed, they mainly address the sim- ple foreground-background segmentation and cannot solve the class-heterogeneous problem for semantic segmentation with a variety of object classes. A recent FL approach, Fed- Drive [14], evaluates FL methods on an autonomous driv- ing semantic segmentation dataset, Cityscapes [9]. How- ever, FedDrive [14] focuses on domain heterogeneity (im- ages from different cities) while ignoring the more challeng- ing class-heterogeneous problem. In this paper, we focus on class-heterogeneous feder- ated learning for semantic segmentation, which has spe- cific and more severe issues compared with classifica- tion. First, images for semantic segmentation are more complex, and pixel-level annotation is extremely time- consuming. Clients usually annotate the objects of fre- quent classes and ignore the rare ones. Due to the non- IID (non-Independent Identically Distribution) data distri- bution of different clients, classes ignored by one client may be foreground classes in another client. For exam- ple, in Fig. 1 (a), the ignored class “person” in Client 1 is annotated in Client 2. The foreground-background incon- sistency across clients leads to divisive optimization direc- tions and degrades the capability of the aggregated global model. Second, as shown in Fig. 1 (b), even if there is no foreground-background inconsistency, for non-IID dis- tribution, complex dense prediction makes the local opti- mization direction diverging farther to the global optimum compared with classification tasks, resulting in poor conver- gence. From the perspective of the pixel embedding space, the local update in each client cannot learn the relative posi- tions of different semantic classes in the pixel embedding space, leading to the confounded embedding space after global aggregation. In this paper, we propose a new federated learning method for semantic segmentation, FedSeg, to address the above issues. A standard objective function for semantic segmentation is the cross-entropy (CE) loss which takes effect on foreground pixels and ignores the background pixels. For FL with non-IID data distribution, it makes the learned local optimum away from the global optimum. Thus, we propose a simple but strong baseline, a modified cross-entropy loss, by aggregating the probabilities of back- ground classes. The modified loss corrects “client drift” in local updates and alleviates the foreground-background in- consistency problem. Then we further introduce a local- to-global pixel-level contrastive learning loss to enforce the local pixel embedding space close to the global semantic space, improving the convergence of the global model. Extensive experiments on four semantic segmentation datasets (Cityscapes [9], CamVID [3], PascalVOC [13] and ADE20k [63]) are conducted to evaluate the effectiveness of our FedSeg. Experimental results show that the sim-ple modified cross-entropy loss significantly improves the segmentation quality. Based on it, our proposed local-to- global pixel contrastive learning consistently improves the segmentation performance compared with previous FL al- gorithms [1, 22, 26, 31]. To summarize, the contributions of this paper are as fol- lows: •We systematically investigate federated learning for the semantic segmentation task with a variety of classes, particularly the class-heterogeneous problem. •We propose a strong baseline with a simple modified CE loss and a local-to-global metrics learning method to alleviate the class distribution drift problem across clients. •We provide benchmarks on four semantic segmenta- tion datasets to evaluate our FedSeg for the semantic seg- mentation FL problem. We hope this work will motivate the FL community to further study the federated learning problem for challenging semantic segmentation tasks.

Li_RiDDLE_Reversible_and_Diversified_De-Identification_With_Latent_Encryptor_CVPR_2023

Abstract This work presents RiDDLE, short for Reversible and Diversified De-identification with Latent Encryptor, to pro- tect the identity information of people from being misused. Built upon a pre-learned StyleGAN2 generator, RiDDLE manages to encrypt and decrypt the facial identity within the latent space. The design of RiDDLE has three appealing properties. First, the encryption process is cipher-guided and hence allows diverse anonymization using different passwords. Second, the true identity can only be de- crypted with the correct password, otherwise the system will produce another de-identified face to maintain the privacy. Third, both encryption and decryption share an efficient implementation, benefiting from a carefully tailored lightweight encryptor. Comparisons with existing alternatives confirm that our approach accomplishes the de-identification task with better quality, higher diversity, and stronger reversibility. We further demonstrate the effectiveness of RiDDLE in anonymizing videos. Code is available in https://github.com/ldz666666/RiDDLE.

1. Introduction Recent advances in deep learning and computer vision technology bring convenience together with security con- cerns. Personal images shared on social media can be collected and abused by unauthorized software and mali- cious attackers, posing a threat to the privacy of individuals. Comparing with all other biometrics, face has unique im- portance because of its extensive application scenarios and abundant personal information. Face de-identification aims to hide the identity in the face image or video stream for privacy protection. Traditional de-identification methods such as blurring and mosaicing can effectively obfuscate the identity information, but the protected image is often *Corresponding author. Ori Encrypted1 Encrypted2 DecryptedWrongly Decrypted1 Wrongly Decrypted2 Figure 1. Encryption and decryption on images in the wild. The first column shows the original people. The second and third columns show different encrypted faces according to different passwords. The fourth column is the correctly decrypted faces, and the last two columns are the incorrectly decrypted faces. severely damaged and loses its utility. Current de-identification methods can generate a similar- looking person with a changed identity and have improved the image quality and utility by a large margin. How- ever, many existing works [6, 10, 25] tend to generate faces with homogeneous appearances for different people, which is easy for an hacker to realize that these faces are anonymized. In some cases, e.g. online conferences [17], the de-identified faces of participants should be different from each other. For reliable identity protection, it is also necessary to consider variations in ethnicity, age, gender and other facial features. Therefore, diversity is important to face identification. Meanwhile, it is vital for the anonymous faces to keep superior image quality and utility. The former brings better photo-realism and stronger safety, while the latter makes it possible to perform identity- agnostic task on an anonymous face image in a privacy- preserving manner. Moreover, most of the previous works only pay atten- tion to the process of de-identification, and neglect the importance of identity recovery. Reversibility of the 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. 8093 identification system is also crucial. For instance, family members are more willing to see the real face rather than the de-identified one and people may want to share data that only certain authorized parties can interpret. Generally, a well-developed de-identification model should hold the following properties. a) Maintaining high quality and the utility of the anonymous faces, as well as the identity independent attributes. b) Generating diversified virtual identities for safer privacy protection. c) Recovering the original faces when security conditions are satisfied. To achieve the above goals, we propose RiDDLE, which is short for Revers ible and Diversified De-identification with Latent Encryptor. The main features of our framework are shown below. Better Quality. First, we project the face image onto the latent space of a strong generator, StyleGAN2 [13]. Due to the decoupling characteristics of the face manifold, the identity independent attributes can be largely preserved. At the same time, high-quality virtual faces can be synthesized. De-identification and recovery results on images in the wild are shown in Figure 1. Higher Diversity. After obtaining the inverted la- tent code, we perform encryption and decryption with a lightweight latent encryptor together with several randomly sampled passwords. In the encryption phase, each password is associated with a unique identity. Discrepancy between different anonymous faces is maximized, resulting in high diversity. Stronger Reversibility. In the decryption phase, when the password is correct, the true identity can be restored. Otherwise, a new de-identified face with photo-realism is returned. Different from opponents [7] and [4] which can achieve identity recovery to some extent, RiDDLE is free from manually designed encryption rules, does not need to be retrained for different passwords, and brings fewer visual artifacts. Another advantage of performing latent encryption is that when face datasets are unavailable due to privacy reasons, randomly sampled latents can be used to train the encryptor. We evaluate RiDDLE on both face image and video de-identification tasks. Sufficient experiments on public datasets and in the wild data have proven the superiority of RiDDLE over previous works.

Qi_Real-Time_6K_Image_Rescaling_With_Rate-Distortion_Optimization_CVPR_2023

Abstract Contemporary image rescaling aims at embedding a high-resolution (HR) image into a low-resolution (LR) thumbnail image that contains embedded information for HR image reconstruction. Unlike traditional image super- resolution, this enables high-fidelity HR image restoration faithful to the original one, given the embedded informa- tion in the LR thumbnail. However, state-of-the-art image rescaling methods do not optimize the LR image file size for efficient sharing and fall short of real-time performance for ultra-high-resolution ( e.g., 6K) image reconstruction. To address these two challenges, we propose a novel frame- work (HyperThumbnail) for real-time 6K rate-distortion- aware image rescaling. Our framework first embeds an HR image into a JPEG LR thumbnail by an encoder with our proposed quantization prediction module, which minimizes the file size of the embedding LR JPEG thumbnail while maximizing HR reconstruction quality. Then, an efficient frequency-aware decoder reconstructs a high-fidelity HR image from the LR one in real time. Extensive experiments demonstrate that our framework outperforms previous im- age rescaling baselines in rate-distortion performance and can perform 6K image reconstruction in real time.

1. Introduction With an increasing number of high-resolution (HR) im- ages being produced and shared by users on the internet, a new challenge has arisen: how can we store and transfer HR images efficiently? Storing HR images on the cloud, such as iCloud, is becoming a widely adopted solution that saves storage on a user’s mobile device ( e.g., smartphones) as only their low-resolution (LR) counterparts are stored on the mobile device for an instant preview. However, when a user wants to obtain the full-resolution image, the entire HR image must be downloaded on the fly from the cloud, which can result in a poor user experience when the internet connection is unstable or not available. Real-time image rescaling can serve as a competitive so- lution to improving the user experience of cloud photo stor- *Equal contribution. Restored HR Timeout✘ PSNR: 34.20 Cloud StorageReal-timeDecoderEncoder 1440x810, 1.37MB 5760x3240, 48.3MB Downsample and compressDecompress and upsampleEncoder Cloud Server HR HyperThumbnailLRFigure 1. The application of 6K image rescaling in the con- text of cloud photo storage on smartphones ( e.g., iCloud). As more high-resolution (HR) images are uploaded to cloud stor- age nowadays, challenges are brought to cloud service providers (CSPs) in fulfilling latency-sensitive image reading requests ( e.g., zoom-in) through the internet. To facilitate faster transmission and high-quality visual content, our HyperThumbnail framework helps CSPs to encode an HR image into an LR JPEG thumbnail, which users could cache locally. When the internet is unstable or un- available, our method can still reconstruct a high-fidelity HR im- age from the JPEG thumbnail in real time. age, as shown in Fig. 1. Such a solution can first embed an HR image (on the cloud) into an LR JPEG thumbnail (on the mobile device) by an encoder, and the thumbnail pro- vides an instant preview with little storage. When the user wants to zoom in on the thumbnail, the HR image with fine details can be reconstructed locally in real time. In addition, image rescaling has other applications in image sharing, as it can “bypass” the resolution limitation of some platforms (e.g., WhatsApp) to reconstruct a high-quality HR image from an LR one [57]. While modern smartphones and cam- eras can capture ultra-high-resolution images in 4K (iPhone 13) or even 6K (Blackmagic camera), we are interested in designing a real-time image rescaling framework for ultra- high-resolution images ( e.g., 4K or 6K), which minimizes LR file size while maximizing HR and LR image quality. However, existing image rescaling methods have their own flaws in practice, as shown in Table 1 where we compare different image rescaling methods in terms of their properties. One potential solution is to upsample the downsampled thumbnail with super-resolution (SR) meth- 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. 14092 Method(a) Downsampled JPEG + super-resolution [37](b) Flow-based rescaling [36, 55] (c) Ours Architecture HRLR.jpg𝐻𝑅% HR𝐻𝑅#LR HR𝐻𝑅#LR.jpg Reconstruction fidelity % ! ! Rate-distortion optimization % % ! Real-time 6K reconstruction – % ! Table 1. The comparison of different methods related to image rescaling. (a) Super-resolution from downsampled JPEG does not optimize rate-distortion performance and can hardly maintain high fidelity due to information lost in downsampling. (b) SOTA flow-based image rescaling methods also ignore the file size constraints and are not real-time for 6K reconstruction due to the limited speed of invertible networks. (c) Our framework optimizes rate-distortion performance while maintaining high-fidelity and real-time 6K image rescaling. ods [14, 17, 35, 37, 61, 62] (Table 1(a)). However, such a framework applies a simple downsampling strategy ( e.g., Bilinear, Bicubic) to the HR image so that high-frequency details are basically lost in the LR thumbnail. Also, SR methods only focus on HR reconstruction, which leads to a sub-optimal image rescaling performance. Instead, ded- icated image rescaling approaches aim to embed informa- tion into a visually pleasing LR image and then recon- struct the HR image with an upsampling module. Recently, state-of-the-art image rescaling works utilize normalizing flow [25,36,55,57] show impressive image embedding and reconstruction capability that outperforms SR approaches, in terms of the reconstructed HR image quality. However, there are still some great challenges to apply these flow- based rescaling frameworks in real-world applications, as shown in Table 1(b). First, the file size of the LR thumbnail is not optimized. Second, the reconstruction stage of these image rescaling methods is computationally expensive due to their invertible network architecture with extensive use of dense blocks [23]: IRN [55] costs about a second to re- construct a 4K image with 4x rescaling on a modern GPU, which is far from real time (Table 2). In this work, we propose the HyperThumbnail , a rate- distortion-aware framework for 6K real-time image rescal- ing, as shown in Table 1(c). In this framework, we embed an HR image into a low-bitrate JPEG thumbnail by an en- coder and a quantization table predictor, as JPEG is a dom- inant image compression format today [3]. Then the JPEG thumbnail can be upscaled to its high-fidelity HR counter- part with our efficient decoder in real time. We leverage an asymmetric encoder-decoder architecture, where most computation is put in the encoder to keep the decoder lightweight. This makes it possible for our decoder to up- scale a thumbnail to 6K in real time, significantly faster than previous flow-based image rescaling methods [36, 55]. Meanwhile, the Rate-Distortion (RD) performance is an important and practical metric rarely studied in prior rescal- ing works. In this paper, we define the rate as the ratio between the thumbnail file size and the number of pixelsin the HR image, also known as the bits-per-pixel (bpp). The distortion consists of two parts: the perceptual qual- ity of the thumbnail (LR distortion) and the fidelity of the restored HR image (HR distortion). The rate-distortion per- formance evaluates an image rescaling framework in both storage cost and visual quality. Without explicit RD con- straints, recent works in image rescaling [36, 55] do not consider RD performance in their models. While some works [26,47,53,54,58] leverage the rate constraint by em- bedding extra information in JPEG, they simply utilize a fixed differentiable JPEG module, which we argue is sub- optimal for image rescaling. Because such a process dete- riorates the information in the embedding images without considering their local distribution. Moreover, the quanti- zation process of JPEG introduces noise in the frequency domain and introduces well-known JPEG artifacts, which brings great challenges to information restoration. To remedy these issues, our image rescaling framework is designed to jointly optimize image quality and bpp with entropy models. Instead of using fixed quantization ta- bles in conventional JPEG (Sec. 3.1), we propose a novel quantization prediction module (QPM) that predicts image- adaptive quantization tables, which can optimize RD perfor- mance. We further adopt a frequency-aware decoder which alleviates JPEG artifacts in the thumbnails and improves HR reconstruction. Moreover, our asymmetric encoder- decoder framework can be extended to optimization-based compression. Our contributions are summarized as follows: • We propose a 6K real-time rescaling framework with an asymmetric encoder-decoder architecture, named HyperThumbnail, which embeds a high-resolution im- age into a JPEG thumbnail that can be viewed in pop- ular browsers. The decoder utilizes both spatial and frequency information to reconstruct high-fidelity im- ages in real time for 6K image upsampling. • We introduce a new quantization prediction module (QPM) that improves the RD performance in the en- 14093 coding stage of our framework. Furthermore, we adopt rate-distortion-aware loss functions along with QPM to optimize the RD performance. • Experiments show that our framework outperforms state-of-the-art image rescaling methods with higher LR and HR image quality and faster reconstruction speed at similar file size.

Olson_Cross-GAN_Auditing_Unsupervised_Identification_of_Attribute_Level_Similarities_and_Differences_CVPR_2023

Abstract Generative Adversarial Networks (GANs) are notori- ously difficult to train especially for complex distributions and with limited data. This has driven the need for tools to audit trained networks in human intelligible format, for example, to identify biases or ensure fairness. Existing GAN audit tools are restricted to coarse-grained, model- data comparisons based on summary statistics such as FID or recall. In this paper, we propose an alternative ap- proach that compares a newly developed GAN against a prior baseline. To this end, we introduce Cross-GAN Audit- ing(xGA) that, given an established “reference” GAN and a newly proposed “client” GAN, jointly identifies intelligi- ble attributes that are either common across both GANs, novel to the client GAN, or missing from the client GAN.This provides both users and model developers an intuitive assessment of similarity and differences between GANs. We introduce novel metrics to evaluate attribute-based GAN auditing approaches and use these metrics to demonstrate quantitatively that xGA outperforms baseline approaches. We also include qualitative results that illustrate the com- mon, novel and missing attributes identified by xGA from GANs trained on a variety of image datasets1.

1. Introduction Generative Adversarial Networks (GANs) [12, 19–21] have become ubiquitous in a range of high impact commer- cial and scientific applications [5, 7–9, 13]. With this pro- 1Source code is available at https : / / github . com / mattolson93/cross_gan_auditing 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. 7981 lific use comes a growing need for investigative tools that are able to evaluate, characterize and differentiate one GAN model from another, especially since such differences can arise from a wide range of factors – biases in training data, model architectures and hyper parameters used in training etc. In practice, this has been mostly restricted to compar- ing two or more GAN models against the dataset they were trained on using summary metrics such as Fr ´echet Inception Distance (FID) [16] and precision/recall [20] scores. However, in many real world scenarios, different models may not even be trained on the same dataset, thereby mak- ing such summary metrics incomparable. More formally, if we define the model comparison problem as one being be- tween a known – and presumably well vetted – reference GAN and a newly developed client GAN. For example, the reference GANs can correspond to models purchased from public market places such as AWS [2], Azure [3], or GCP [11], or to community-wide standards. Furthermore, there is a critical need for more fine-grained, interpretable, investigative tools in the context of fairness and account- ability. Broadly, these class of methods can be studied un- der the umbrella of AI model auditing [1, 6, 32]. Here, the interpretability is used in the context to indicate that the proposed auditing result will involves of human intelligi- ble attributes, rather than summary statistic that do not have explicit association with meaningful semantics. While auditing classifiers has received much attention in the past [32], GAN auditing is still a relatively new research problem with existing efforts focusing on model-data com- parisons, such as identifying how faithfully a GAN recovers the original data distribution [1]. In contrast, we are inter- ested in developing a more general framework that enables a user to visually audit a “client” GAN model with respect the “reference”. This framework is expected to support different kinds of auditing tasks: (i) comparing different GAN models trained on the same dataset (e.g. StyleGAN3- Rotation and StyleGAN3-Translate on FFHQ); (ii) compar- ing models trained on datasets with different biases (e.g., StyleGAN with race imbalance vs StyleGAN with age im- balance); and finally (iii) comparing models trained using datasets that contain challenging distribution shifts (e.g., CelebA vs Toons). Since these tools are primarily intended for human experts and auditors, interpretability is critical. Hence, it is natural to perform auditing in terms of human intelligible attributes. Though there has been encouraging progress in automatically discovering such attributes from a single GAN in the recent years [14, 28, 39, 40, 43] they are not applicable to our setting with multiple GANs. Proposed work We introduce cross-GAN auditing (xGA), an unsupervised approach for identifying attribute similar- ities and differences between client GANs and reference models (which could be pre-trained and potentially unre- lated). Since the GANs are trained independently, their la-tent spaces are disparate and encode different attributes, and thus they are not directly comparable. Consequently, dis- covering attributes is only one part of the solution; we also need to ‘align’ humanly meaningful and commonly occur- ring attributes across the individual latent spaces. Our audit identifies three distinct sets of attributes: (a) common: attributes that exist in both client and refer- ence models; (b) novel: attributes encoded only in the client model; (c) missing: attributes present only in the reference. In order to identify common attributes, xGA exploits the fact that shared attributes should induce similar changes in the resulting images across both the models. On the other hand, to discover novel/missing attributes, xGA leverages the key insight that attribute manipulations unique to one GAN can be viewed as out of distribution (OOD) to the other GAN. Using empirical studies with a variety of Style- GAN models and benchmark datasets, we demonstrate that xGA is effective in providing a fine-grained characterization of generative models. Contributions (i) We present the first cross-GAN audit- ing framework that uses an unified, attribute-centric method to automatically discover common, novel, and missing at- tributes from two or more GANs; (ii) Using an external, robust feature space for optimization, xGA produces high- quality attributes and achieves effective alignment even across challenging distribution shifts; (iii) We introduce novel metrics to evaluate attribute-based GAN auditing ap- proaches; and (iv) We evaluate xGA using StyleGANs trained on CelebA, AFHQ, FFHQ, Toons, Disney and Met- Faces, and also provide a suite of controlled experiments to evaluate cross-GAN auditing methods.

Li_Towards_High-Quality_and_Efficient_Video_Super-Resolution_via_Spatial-Temporal_Data_Overfitting_CVPR_2023

Abstract As deep convolutional neural networks (DNNs) are widely used in various fields of computer vision, leveraging the overfitting ability of the DNN to achieve video resolu- tion upscaling has become a new trend in the modern video delivery system. By dividing videos into chunks and over- fitting each chunk with a super-resolution model, the server encodes videos before transmitting them to the clients, thus achieving better video quality and transmission efficiency. However, a large number of chunks are expected to ensure good overfitting quality, which substantially increases the storage and consumes more bandwidth resources for data transmission. On the other hand, decreasing the number of chunks through training optimization techniques usually re- quires high model capacity, which significantly slows down execution speed. To reconcile such, we propose a novel method for high-quality and efficient video resolution up- scaling tasks, which leverages the spatial-temporal infor- mation to accurately divide video into chunks, thus keep- ing the number of chunks as well as the model size to min- imum. Additionally, we advance our method into a sin- gle overfitting model by a data-aware joint training tech- nique, which further reduces the storage requirement with negligible quality drop. We deploy our models on an off- the-shelf mobile phone, and experimental results show that our method achieves real-time video super-resolution with high video quality. Compared with the state-of-the-art, our method achieves 28 fps streaming speed with 41.6 PSNR, which is 14 ×faster and 2.29 dB better in the live video resolution upscaling tasks. Code available in https:// github.com/coulsonlee/STDO-CVPR2023.git .

1. Introduction Being praised by its high image quality performance and wide application scenarios, deep learning-based super- resolution (SR) becomes the core enabler of many incred- †Equal Contribution. Frame: 250/375 Frame: 350/375 Figure 1. Patch PSNR heatmap of two frames in a 15s video when super-resolved by a general WDSR model. A clear bound- ary shows that PSNR is strongly related to video content. ible, cutting-edge applications in the field of image/video reparation [10, 11, 39, 40], surveillance system enhance- ment [9], medical image processing [35], and high-quality video live streaming [20]. Distinct from the traditional methods that adopt classic interpolation algorithms [15, 45] to improve the image/video quality, the deep learning-based approaches [10, 11, 21, 24, 28, 40, 44, 47, 57, 60] exploit the advantages of learning a mapping function from low- resolution (LR) to high-resolution (HR) using external data, thus achieving better performance due to better generaliza- tion ability when meeting new data. Such benefits have driven numerous interests in design- ing new methods [5, 17, 50] to deliver high-quality video stream to users in the real-time fashion, especially in the context of massive online video and live streaming avail- able. Among this huge family, an emerging representa- tive [13,16,31,38] studies the prospect of utilizing SR model to upscale the resolution of the LR video in lieu of transmit- ting the HR video directly, which in many cases, consumes tremendous bandwidth between servers and clients [19]. One practical method is to deploy a pretrained SR model on the devices of the end users [25, 54], and perform res- olution upscaling for the transmitted LR videos, thus ob- taining HR videos without causing bandwidth congestion. However, the deployed SR model that is trained with lim- ited data usually suffers from limited generalization abil- 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. 10259 Clustering function SR model 2SR model 1LR video HR video ... ... ...... ... Chunk1 (low PSNR patches) Chunk2 (high PSNR patches)Train Train Video framesSlice Ungrouped patchesGrouped patchesEvaluate Deliver Overfitting each video chunkTraining Super-resolution at user endServer process Overfit and transmit User end AI Frame: 350/375Frame: 280/375Frame: 210/375 Frame: 350/375Frame: 280/375Frame: 210/375 & rankFigure 2. Overview of the proposed STDO method. Each video frame is sliced into patches, and all patches across time dimension are divided and grouped into chunks. Here we set the number of chunks to 2 for clear illustration. Then each chunk is overfitted by independent SR models, and delivered to end-user for video super-resolution. ity, and may not achieve good performance at the presence of new data distribution [55]. To overcome this limitation, new approaches [4, 8, 20, 30, 51, 53, 55] exploit the overfit- ting property of DNN by training an SR model for each video chunk (i.e., a fragment of the video), and deliver- ing the video alongside the corresponding SR models to the clients. This trade-off between model expressive power and the storage efficiency significantly improves the quality of the resolution upscaled videos. However, to obtain bet- ter overfitting quality, more video segments are expected, which notably increase the data volume as well as system overhead when processing the LR videos [55]. While ad- vanced training techniques are proposed to reduce the num- ber of SR models [30], it still requires overparameterized SR backbones (e.g., EDSR [28]) and handcrafted modules to ensure sufficient model capacity for the learning tasks, which degrades the execution speed at user-end when the device is resource-constraint. In this work, we present a novel approach towards high- quality and efficient video resolution upscaling via Spatial- Temporal DataOverfitting, namely STDO , which for the first time, utilizes the spatial-temporal information to accu- rately divide video into chunks. Inspired by the work pro- posed in [1, 14, 23, 46, 58] that images may have different levels of intra- and inter-image (i.e., within one image or between different images) information density due to var- ied texture complexity, we argue that the unbalanced infor- mation density within or between frames of the video uni- versally exists, and should be properly managed for data overfitting. Our preliminary experiment in Figure 1 showsthat the PSNR values at different locations in a video frame forms certain pattern regarding the video content, and ex- hibits different patterns along the timeline. Specifically, at the server end, each frame of the video is evenly divided into patches, and then we split all the patches into multi- ple chunks by PSNR regarding all frames. Independent SR models will be used to overfit the video chunks, and then de- livered to the clients. Figure 2 demonstrates the overview of our proposed method. By using spatial-temporal informa- tion for data overfitting, we reduce the number of chunks as well as the overfitting models since they are bounded by the nature of the content, which means our method can keep a minimum number of chunks regardless the dura- tion of videos. In addition, since each chunk has similar data patches, we can actually use smaller SR model without handcrafted modules for the overfitting task, which reduces the computation burden for devices of the end-user. Our experimental results demonstrate that our method achieves real-time video resolution upscaling from 270p to 1080p on an off-the-shelf mobile phone with high PSNR. Note that STDO encodes different video chunks with independent SR models, we further improve it by a Joint training technique ( JSTDO ) that results in one single SR model for all chunks, which further reduces the storage requirement. We design a novel data-aware joint training technique, which trains a single SR model with more data from higher information density chunks and less data from their counterparts. The underlying rationale is consistent with the discovery in [46, 58], that more informative data contributes majorly to the model training. We summarize 10260 our contributions as follows: • We discover the unbalanced information density within video frames, and it universally exists and constantly changes along the video timeline. • By leveraging the unbalanced information density in the video, we propose a spatial-temporal data overfit- ting method STDO for video resolution upscaling, which achieves outperforming video quality as well as real-time execution speed. • We propose an advanced data-aware joint training tech- nique which takes different chunk information density into consideration, and reduces the number of SR mod- els to a single model with negligible quality degradation. • We deploy our models on an off-the-shelf mobile phone, and achieve real-time super-resolution performance.

Miangoleh_Realistic_Saliency_Guided_Image_Enhancement_CVPR_2023

Abstract Common editing operations performed by profes- sional photographers include the cleanup operations: de- emphasizing distracting elements and enhancing subjects. These edits are challenging, requiring a delicate balance between manipulating the viewer’s attention while main- taining photo realism. While recent approaches can boast successful examples of attention attenuation or amplifica- tion, most of them also suffer from frequent unrealistic ed- its. We propose a realism loss for saliency-guided image en- hancement to maintain high realism across varying image types, while attenuating distractors and amplifying objects of interest. Evaluations with professional photographers confirm that we achieve the dual objective of realism and ef- fectiveness, and outperform the recent approaches on their own datasets, while requiring a smaller memory footprint and runtime. We thus offer a viable solution for automating image enhancement and photo cleanup operations.

1. Introduction In everyday photography, the composition of a photo typically encompasses subjects on which the photographer intends to focus our attention, rather than other distracting things. When distracting things cannot be avoided, photog- raphers routinely edit their photos to de-emphasize them. Conversely, when the subjects are not sufficiently visible, photographers routinely emphasize them. Among the most common emphasis and de-emphasis operations performed by professionals are the elementary ones: changing the sat- uration, exposure, or the color of each element. Although conceptually simple, these operations are challenging to ap- ply because they must delicately balance the effects on the viewer attention with photo realism. To automate this editing process, recent works use saliency models as a guide [1,2,4,8,16,17]. These saliency models [3, 7, 10, 14, 19] aim to predict the regions in 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. 186 image that catch the viewer’s attention, and saliency-guided image editing methods are optimized to increase or decrease the predicted saliency of a selected region. Optimizing solely based on the predicted saliency, however, often re- sults in unrealistic edits, as illustrated in Fig. 1. This issue results from the instability of saliency models under the im- age editing operations, as saliency models are trained on unedited images. Unrealistic edits can have low predicted saliency even when they are highly noticeable to human ob- servers, or vice versa. This was also noted by Aberman et al. [1], and is illustrated in Fig. 2. Previous methods tried to enforce realism using adver- sarial setups [2,4,8,17], GAN priors [1,8], or cycle consis- tency [2] but with limited success (Fig. 1). Finding the exact point when an image edit stops looking realistic is challeng- ing. Rather than focusing on the entire image, in this work, we propose a method for measuring the realism of a local edit. To train our network, we generate realistic image ed- its by subtle perturbations to exposure, saturation, color or white balance, as well as very unrealistic edits by apply- ing extreme adjustments. Although our network is trained with only positive and negative examples at the extremes, we successfully learn a continuous measure of realism for a variety of editing operations as shown in Fig. 3. We apply our realism metric to saliency-guided image editing by training the system to optimize the saliency of a selected region while being penalized for deviations from realism. We show that a combined loss allows us to enhance or suppress a selected region successfully while maintaining high realism. Our method can be also be applied to multiple regions in a photograph as shown in Fig. 1. Evaluations with professional photographers and photo editors confirm our claim that we maintain high realism and succeed at redirecting attention in the edited photo. Further, our results are robust to different types of images including human faces, and are stable across different permutations of edit parameters. Taken together with our model size of 26Mb and run-time of 8ms, these results demonstrate that we have a more viable solution for broader use than the ap- proaches that are available for these tasks to date.

Long_PointClustering_Unsupervised_Point_Cloud_Pre-Training_Using_Transformation_Invariance_in_Clustering_CVPR_2023

Abstract Feature invariance under different data transformations, i.e., transformation invariance, can be regarded as a type of self-supervision for representation learning. In this pa- per, we present PointClustering, a new unsupervised rep- resentation learning scheme that leverages transformation invariance for point cloud pre-training. PointClustering formulates the pretext task as deep clustering and employs transformation invariance as an inductive bias, following the philosophy that common point cloud transformation will not change the geometric properties and semantics. Techni- cally, PointClustering iteratively optimizes the feature clus- ters and backbone, and delves into the transformation in- variance as learning regularization from two perspectives: point level and instance level. Point-level invariance learn- ing maintains local geometric properties through gathering point features of one instance across transformations, while instance-level invariance learning further measures cluster- s over the entire dataset to explore semantics of instances. Our PointClustering is architecture-agnostic and readily applicable to MLP-based, CNN-based and Transformer- based backbones. We empirically demonstrate that the models pre-learnt on the ScanNet dataset by PointClus- tering provide superior performances on six benchmark- s, across downstream tasks of classification and segmen- tation. More remarkably, PointClustering achieves an ac- curacy of 94.5% on ModelNet40 with Transformer back- bone. Source code is available at https://github. com/FuchenUSTC/PointClustering .

1. Introduction 3D point cloud analysis has seen tremendous progress and made great success in industrial applications, e.g., au- tonomous driving, augmented reality and robotics. The achievements heavily rely on large quantities of human an- Dinstance feature Clustering Learning1 32 4 1 2 transformation transformationpoint feature  AB transformation C DClustering LearningCView 1 View 2Scene 1 Scene 1 Scene 2(b) point level invariance learningpoint cloudsback‐propagation backbone point cloudstransformationView 1 View 2ClusteringDistance  Optimization (a) clustering learning on point cloud 3 4 (c) instance level invariance learningView 1 View 2Figure 1. Illustration of (a) clustering learning on point cloud by using feature invariance at (b) point level and (c) instance level. notations for supervised learning. However, acquiring and manual labeling 3D point cloud data is very expensive and time-consuming, while the underlying rich data structure is also not yet fully leveraged. In contrast, unsupervised learn- ing leaves it on its own to characterize the underlying fea- ture distribution completely on data itself and is therefore an appealing way towards more generic model pre-training. The research in unsupervised point cloud pre-training has mainly proceeded along two dimensions with respec- t to the formulation of pretext task: contrastive learning [23, 65, 74] and reconstruction [34, 52, 60]. Early works of contrastive learning generally suggest to leverage point or scene discrimination across different views [65] or modal- ities [1, 74] for similarity learning. Instead, the direction of point cloud reconstruction [34, 60] formulates the learn- ing target as shape completion from the partial points. Un- like existing discrimination or reconstruction paradigm in 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. 21824 sample-specific manner, clustering technique estimates the data distribution holistically for class level . We rely on such recipe and shape a new unsupervised point cloud pre- training scheme that capitalizes on deep clustering as the pretext task. Technically, we iteratively optimize feature clusters and backbone as shown in Figure 1(a), and utilize transformation invariance as an inductive bias. We look into the feature invariance learning across data transformations from two aspects: point level and instance level. The ratio- nale behind point level feature invariance is that the point features of an identical object (e.g., points of the chair in Figure 1(b)) should be invariant across different transfor- mations since the geometric properties will not change with transformations. Similar in spirit, the high-level semantics of instances across 3D scenes (e.g., the instances of chair in Figure 1(c)) do not vary along with the transformations. As such, we delve into both point-level and instance-level transformation invariance to regulate deep clustering. By materializing the idea of transformation invariance as regularization for deep clustering, we present a novel PointClustering approach for unsupervised point cloud pre- training. Specifically, we first obtain the instance masks of objects in each 3D scene via Density-Based Spatial Cluster- ing of Applications with Noise (DBSCAN) [12] algorithm. Based on the instance masks, the point features of an iden- tical object under different transformations are clustered to- gether to characterize geometric properties of points. The instance-level feature of one object is then computed by globally pooling all point features of that object. Given al- l instance features over the entire dataset, PointClustering further seeks the feature consistency across transformations at instance level. We employ InfoNCE loss to optimize the similarity between points or instances and their correspond- ing clustering centroids (i.e., prototypes). The main contribution of this work is a new paradigm that leverages feature invariance under different data trans- formations for unsupervised point cloud pre-training. The solution also leads to the elegant view of how to explore self-supervision from the standpoint of transformation in- variance, and how to indicate geometric properties and se- mantics of point cloud for unsupervised pre-training. Ex- tensive experiments on six benchmarks over three down- stream tasks verify that PointClustering outperforms the state-of-the-art unsupervised pre-training models.

Prabhu_Computationally_Budgeted_Continual_Learning_What_Does_Matter_CVPR_2023

Abstract Continual Learning (CL) aims to sequentially train mod- els on streams of incoming data that vary in distribution by preserving previous knowledge while adapting to new data. Current CL literature focuses on restricted access to previously seen data, while imposing no constraints on the computational budget for training. This is unreasonable for applications in-the-wild, where systems are primarily constrained by computational and time budgets, not stor- age. We revisit this problem with a large-scale benchmark and analyze the performance of traditional CL approaches in a compute-constrained setting, where effective memory samples used in training can be implicitly restricted as a consequence of limited computation. We conduct experi- ments evaluating various CL sampling strategies, distillation losses, and partial fine-tuning on two large-scale datasets, namely ImageNet2K and Continual Google Landmarks V2 in data incremental, class incremental, and time incremen- tal settings. Through extensive experiments amounting to a total of over 1500 GPU-hours, we find that, under compute- constrained setting, traditional CL approaches, with no ex- ception, fail to outperform a simple minimal baseline that samples uniformly from memory. Our conclusions are con- sistent in a different number of stream time steps, e.g., 20 to 200, and under several computational budgets. This suggests that most existing CL methods are particularly too computationally expensive for realistic budgeted de- ployment. Code for this project is available at: https: //github.com/drimpossible/BudgetCL .

1. Introduction Deep learning has excelled in various computer vision tasks [8,21,25,43] by performing hundreds of shuffled passes through well-curated offline static labeled datasets. However, modern real-world systems, e.g., Instagram, TikTok, and Flickr, experience high throughput of a constantly changing stream of data, which poses a challenge for deep learning to cope with such a setting. Continual learning (CL) aims to go beyond static datasets and develop learning strategies that can adapt and learn from streams where data is pre- *authors contributed equally; order decided by a coin flip. Figure 1. Main Findings. Under per time step computationally budgeted continual learning, classical continual learning methods, e.g., sampling strategies, distillation losses, and fully connected (FC) layer correction based methods such as calibration, struggle to cope with such a setting. Most proposed continual algorithms are particularly useful only when large computation is available, where, otherwise, minimalistic algorithms (ERM) are superior. sented incrementally over time, often referred to as time steps. However, the current CL literature overlooks a key necessity for practical real deployment of such algorithms. In particular, most prior art is focused on offline continual learning [22, 23, 41] where, despite limited access to previ- ous stream data, training algorithms do not have restrictions on the computational training budget per time step. High-throughput streams, e.g., Instagram, where every stream sample at every time step needs to be classified for, say, misinformation or hate speech, are time-sensitive in which long training times before deployment are simply not an option. Otherwise, new stream data will accumulate until training is completed, causing server delays and worsening user experience. Moreover, limiting the computational budget is necessary towards reducing the overall cost. This is because computa- tional costs are higher compared to any storage associated costs. For example, on Google Cloud Standard Storage (2¢per GB per month), it costs no more than 6 ¢to store the entire CLEAR benchmark [26], a recent large-scale CL dataset. On the contrary, one run of a CL algorithm on CLEAR performing ∼300K iterations costs around 100$ on an A100 Google instance (3 $per hour for 1 GPU). There- fore, it is prudent to have computationally budgeted methods where the memory size, as a consequence, is implicitly re- stricted. This is because, under a computational budget, it is no longer possible to revisit all previous data even if 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. 3698 were all stored in memory (given their low memory costs). This raises the question: “ Do existing continual learning algorithms perform well under per step restricted comp

Michaeli_Alias-Free_Convnets_Fractional_Shift_Invariance_via_Polynomial_Activations_CVPR_2023

Abstract Although CNNs are believed to be invariant to transla- tions, recent works have shown this is not the case due to aliasing effects that stem from down-sampling layers. The existing architectural solutions to prevent the aliasing ef- fects are partial since they do not solve those effects that originate in non-linearities. We propose an extended anti- aliasing method that tackles both down-sampling and non- linear layers, thus creating truly alias-free, shift-invariant CNNs1. We show that the presented model is invariant to integer as well as fractional (i.e., sub-pixel) translations, thus outperforming other shift-invariant methods in terms of robustness to adversarial translations.

1. Introduction Convolutional Neural Networks (CNNs) are the most common model in the image classification field. They were originally intended to have two properties: 1. Shift-invariant output: when we spatially translate the input image, their output does not change. 2. Shift-equivariant representation: when we spatially translate the input image, their internal representation translates in the same way. Both these properties are thought to be beneficial for gener- alization ( i.e., they are useful inductive biases), as we expect the image class not to change by an image translation, and its features to shift together with the image. Moreover, with- out the first property, the CNN might become vulnerable to adversarial attacks using image translations. Such attacks are real threats since they are very simple to execute in a “black-box” setting (where we do not know anything about the CNN). For example, consider a person trying to fool a CNN-based face scanner, by simply moving continuously until a face match is achieved. It was commonly assumed that these useful properties were maintained since CNNs use only shift-equivariant op- erations: the convolution operation and component-wise 1Our code is available at github.com/hmichaeli/alias free convnets/.non-linearities. However, CNN models typically also in- clude downsampling operations such as pooling and strided convolution. Unfortunately, these operations violate equiv- ariance, and this also leads to CNNs not being shift- invariant. Specifically, Azulay and Weiss [2] have shown that shifting an input image by even one pixel can cause the output probability of a trained classifier to change signifi- cantly. This vulnerability can be further exploited in adver- sarial attacks, lowering classifiers’ accuracy by more than 20% [8]. Later, Zhang [33] has shown that this problem- atic behavior stems from an aliasing effect, taking place in downsampling operations such as pooling and strided con- volutions, and non-linear operations on the downsampled signals. Previous works have shown an improvement in CNN invariance to translations using partial solutions that re- duced aliasing. For example, Zhang [33] has suggested adding a low-pass filter before the downsampling opera- tions. This approach has been shown to reduce aliasing caused by downsampling, thus improving shift-invariance, as well as accuracy and noise robustness. Karras et al. [17] have addressed aliasing in the generator within generative adversarial networks (GANs). They have shown that with- out proper treatment, aliasing in GANs leads to a decou- pling of the high-frequency features (texture) from the low- frequency content (structure) in the generated images, thus limiting their applicability in smooth video generation. To alleviate this issue, Karras et al . [17] extended the low- pass filter approach and suggested a solution for the implicit aliasing caused by non-linearities. Their method wraps the component-wise non-linear operat

Ning_Trap_Attention_Monocular_Depth_Estimation_With_Manual_Traps_CVPR_2023

Abstract Predicting a high quality depth map from a single im- age is a challenging task, because it exists infinite pos- sibility to project a 2D scene to the corresponding 3D scene. Recently, some studies introduced multi-head at- tention (MHA) modules to perform long-range interaction, which have shown significant progress in regressing the depth maps. The main functions of MHA can be loosely summarized to capture long-distance information and re- port the attention map by the relationship between pixels. However, due to the quadratic complexity of MHA, these methods can not leverage MHA to compute depth features in high resolution with an appropriate computational com- plexity. In this paper, we exploit a depth-wise convolution to obtain long-range information, and propose a novel trap attention, which sets some traps on the extended space for each pixel, and forms the attention mechanism by the fea- ture retention ratio of convolution window, resulting in that the quadratic computational complexity can be converted to linear form. Then we build an encoder-decoder trap depth estimation network, which introduces a vision transformer as the encoder, and uses the trap attention to estimate the depth from single image in the decoder. Extensive experi- mental results demonstrate that our proposed network can outperform the state-of-the-art methods in monocular depth estimation on datasets NYU Depth-v2 and KITTI, with sig- nificantly reduced number of parameters. Code is available at: https://github.com/ICSResearch/TrapAttention.

1. Introduction Depth estimation is a classical problem in computer vi- sion (CV) field and is a fundamental component for vari- ous applications, such as, scene understanding, autonomous driving, and 3D reconstruction. Estimating the depth map from a single RGB image is a challenge, since the same 2D scene can project an infinite number of 3D scenes. There- *Corresponding author 2 1 12 1 1 4 3 322 4 4 3 3 422 11 1122 11 11 44 33 332222 44 44 33 33 442 1 12 1 1 4 3 322 4 4 3 3 42 1 12 1 1 4 3 322 4 4 3 3 4 Depth Depth Encoder Extend Input image Input image Feature 4 32 1 4 32 1 Feature 4 32 1 Set the traps Set the traps Attention map Net 3x3 conv. 0 30 0 0 30 01 11 1 1 11 1 4 4 4 44 4 4 402 22 02 22 000 333000 0001 11 1 4 4 4 4000222 222222Figure 1. Illustration of trap attention for monocular deep esti- mation. Note that trap attention can significantly enhance depth estimation, as evidenced by the clearer depth differences between the table/chairs and the background. fore, the traditional depth estimation methods [27, 28, 33] are often only suitable for predicting low-dimension, sparse distances [27], or known and fixed targets [28], which obvi- ously limits their application scenarios. To overcome these constraints, many studies [1, 8, 9, 20] have employed the deep neural networks to directly obtain high-quality depth maps. However, most of these research focuses on improving the performance of depth estimation networks by designing more complex or large-scale mod- els. Unfortunately, such a line of research would render the depth estimation task a simple model scale problem with- out the trade-off between performance and computational budget. Recently, several practitioners and researchers in monoc- ular depth estimation [3, 17, 45] introduced the multi-head attention (MHA) modules to perform the long-range inter- 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. 5033 Figure 2. Overview of our trap depth estimation network, which includes an encoder and a decoder. TB, TI and BS denote the trap block, trap interpolation and block selection unit, respectively. TB is the basic block of our decoder, which decodes the depth feature from coarse to fine in five stages. TB consists of a depth-wise (DW) convolution layer, a trap attention (TA) unit and a convolution based MLP. The size of decoder depends on an arbitrary channel dimension (denoted as C). “⊕” denotes the addition operation. (a)t1 (b)t2 (c)t3 (d)t4 Figure 3. The curves for trap functions used in trap attention with- out the rounded operation. (a) and (d) are two similar curves. (b) has a higher frequency. (c) has a different initial phase with other curves. action, which have shown considerable progress in regress- ing the depth maps. Representative works of such methods are AdaBin [3] and NeW CRFs [45]. Nevertheless, due to the quadratic computational complexity of MHA, the com-putational complexity of high resolution depth map for Ad- aBin or NeW CRFs is typically expensive, i.e., for an h×w image, its complexity is O(h2w2). To reduce the computational complexity, in this work, we firstly exploit a deep-wise convolution layer to compute the long-distance information and then propose an attention mechanism, called trap attention, which leverages various manual traps to remove some features in extended space, and exploits a 3×3convolution window to compute rela- tionship and attention map. As a result, the quadratic com- putational complexity O(h2w2)can be converted to linear formO(hw). As illustrated by the example in Figure 1, the proposed trap attention is highly effective for depth esti- mation, which can allocate more computational resource to- ward the informative features, i.e., edges of table and chairs, and output a refined depth map from coarse depth map. Based on this attention mechanism, we finally build an encoder-decoder depth estimation network, which intro- duces a vision transformer as the encoder, and uses the trap attention to estimate the depth from single image in the de- coder. We can build our depth estimation network of differ- ent scales according to the depth estimation scene, which can obtain a balance between performance and computa- tional budget. Experimental results show that our depth estimation network outperform previous estimation meth- ods by remarkable margin on two most popular indoor and outdoor datasets, NYU [36] and KITTI [11], respectively. 5034 Specifically, our model can obtain consistent predictions with sharp details on visual representations, and achieve the new state-of-the-art performance in monocular depth esti- mation, with only 35% parameters of the prior state-of-the- art methods. In summary, our main contributions are as follows: •We use a depth-wise convolution to capture long- distance information and introduce an extra attention mechanism to compute the relationships between fea- tures, which is an efficient alternative to MHA, result- ing in that the computational complexity is reduced fromO(h2w2)toO(hw). •We propose a novel attention mechanism that can al- locate more computational resource toward the infor- mative features, called trap attention, which is highly effective for depth estimation. •We build an end-to-end trap network for monocular depth estimation, which can obtain the state-of-the-art performance on NYU and KITTI datasets, with signif- icantly reduced number of parameters.

Luan_High_Fidelity_3D_Hand_Shape_Reconstruction_via_Scalable_Graph_Frequency_CVPR_2023

Abstract Despite the impressive performance obtained by recent single-image hand modeling techniques, they lack the capa- bility to capture sufficient details of the 3D hand mesh. This deficiency greatly limits their applications when high-fidelity hand modeling is required, e.g., personalized hand model- ing. To address this problem, we design a frequency split network to generate 3D hand mesh using different frequency bands in a coarse-to-fine manner. To capture high-frequency personalized details, we transform the 3D mesh into the frequency domain, and propose a novel frequency decom- position loss to supervise each frequency component. By leveraging such a coarse-to-fine scheme, hand details that correspond to the higher frequency domain can be preserved. In addition, the proposed network is scalable, and can stop the inference at any resolution level to accommodate dif- ferent hardware with varying computational powers. To quantitatively evaluate the performance of our method in terms of recovering personalized shape details, we intro- duce a new evaluation metric named Mean Signal-to-Noise Ratio (MSNR) to measure the signal-to-noise ratio of each mesh frequency component. Extensive experiments demon- strate that our approach generates fine-grained details for high-fidelity 3D hand reconstruction, and our evaluation metric is more effective for measuring mesh details com- pared with traditional metrics. The code is available at https://github.com/tyluann/FreqHand .

1. Introduction High-fidelity and personalized 3D hand modeling have seen great demand in 3D games, virtual reality, and the emerging Metaverse, as it brings better user experiences, e.g., users can see their own realistic hands in the virtual space instead of the standard avatar hands. Therefore, it is 0 2 4 6 8 10 12×103 Frequency componentsFigure 1. An exemplar hand mesh of sufficient details and its graph frequency decomposition. The x-axis shows frequency compo- nents from low to high. The y-axis shows the amplitude of each component in the logarithm. At the frequency domain, the signal amplitude generally decreases as the frequency increases. of great importance to reconstruct high-fidelity hand meshes that can adapt to different users and application scenarios. Despite previous successes in 3D hand reconstruction and modeling [3, 6, 7, 16, 22, 40, 44, 46], few existing solutions focus on enriching the details of the reconstructed shape, and most current methods fail to generate consumer-friendly high-fidelity hands. When we treat the hand mesh as graph signals, like most natural signals, the low-frequency compo- nents have larger amplitudes than those of the high-frequency parts, which we can observe in a hand mesh spectrum curve (Fig. 1). Consequently, if we generate the mesh purely in the spatial domain, the signals of different frequencies could be biased, thus the high-frequency information can be eas- ily overwhelmed by its low-frequency counterpart. More- over, the wide usage of compact parametric models, such as MANO [32], has limited the expressiveness of personalized details. Even though MANO can robustly estimate the hand pose and coarse shape, it sacrifices hand details for compact- ness and robustness in the parameterization process, so the detail expression ability of MANO is suppressed. To better model detailed 3D shape information, we trans- form the hand mesh into the graph frequency domain, 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. 16795 design a frequency-based loss function to generate high- fidelity hand mesh in a scalable manner. Supervision in the frequency domain explicitly constrains the signal of a given frequency band from being influenced by other fre- quency bands. Therefore, the high-frequency signals of hand shape will not be suppressed by low-frequency sig- nals despite the amplitude disadvantage. To improve the expressiveness of hand models, we design a new hand model of12,337vertices that extends previous parametric models such as MANO with nonparametric representation for resid- ual adjustments. While the nonparametric residual expresses personalized details, the parametric base ensures the over- all structure of the hand mesh, e.g., reliable estimation of hand pose and 3D shape. Instead of fixing the hand mesh resolution, we design our network architecture in a coarse-to- fine manner with three resolution levels U-net for scalability. Different levels of image features contribute to different levels of detail. Specifically, we use low-level features in high-frequency detail generation and high-level features in low-frequency detail generation. At each resolution level, our network outputs a hand mesh with the corresponding resolution. During inference, the network outputs an increas- ingly higher resolution mesh with more personalized details step-by-step, while the inference process can stop at any one of the three resolution levels. In summary, our contributions include the following. 1.We design a high-fidelity 3D hand model for reconstruct- ing 3D hand shapes from single images. The hand repre- sentation provides detailed expression, and our frequency decomposition loss helps to capture the personalized shape information. 2.To enable computational efficiency, we propose a fre- quency split network architecture to generate high-fidelity hand mesh in a scalable manner with multiple levels of de- tail. During inference, our scalable framework supports budget-aware mesh reconstruction when the computa- tional resources are limited. 3.We propose a new metric to evaluate 3D mesh details. It better captures the signal-to-noise ratio of all frequency bands to evaluate high-fidelity hand meshes. The effec- tiveness of this metric has been validated by extensive experiments. We evaluate our method on the InterHand2.6M dataset [29]. In addition to the proposed evaluation met- rics, we also evaluate mean per joint position error (MPJPE) and mesh Chamfer distance (CD). Compared to MANO and other baselines, our proposed method achieves better results using all three metrics.

Ma_Solving_Oscillation_Problem_in_Post-Training_Quantization_Through_a_Theoretical_Perspective_CVPR_2023

Abstract Post-training quantization (PTQ) is widely regarded as one of the most efficient compression methods practically, benefitting from its data privacy and low computation costs. We argue that an overlooked problem of oscillation is in the PTQ methods. In this paper, we take the initiative to ex- plore and present a theoretical proof to explain why such a problem is essential in PTQ. And then, we try to solve this problem by introducing a principled and generalized frame- work theoretically. In particular, we first formulate the os- cillation in PTQ and prove the problem is caused by the dif- ference in module capacity. To this end, we define the mod- ule capacity (ModCap) under data-dependent and data-free scenarios, where the differentials between adjacent modules are used to measure the degree of oscillation. The prob- lem is then solved by selecting top-k differentials, in which the corresponding modules are jointly optimized and quan- tized. Extensive experiments demonstrate that our method successfully reduces the performance drop and is general- ized to different neural networks and PTQ methods. For example, with 2/4bit ResNet- 50quantization, our method surpasses the previous state-of-the-art method by 1.9%. It becomes more significant on small model quantization, e.g. surpasses BRECQ method by 6.61% on MobileNetV 2×0.5.

1. Introduction Deep Neural Networks (DNNs) have rapidly become a research hotspot in recent years, being applied to various *This work was done when Yuexiao Ma was intern at ByteDance Inc. Code is available at: https://github.com/bytedance/MRECG †Corresponding Author: rrji@xmu.edu.cn OscillationFigure 1. Left: Reconstruction loss distribution of BRECQ [17] on0.5scaled MobileNetV 2quantized to 4/4bit. Loss oscilla- tion in BRECQ during reconstruction see red dashed box. Right: Mixed reconstruction granularity (MRECG) smoothing loss oscil- lation and achieving higher accuracy. scenarios in practice. However, as DNNs evolve, better model performance is usually associated with huge resource consumption from deeper and wider networks [8, 14, 28]. Meanwhile, the research field of neural network compres- sion and acceleration, which aims to deploy models in resource-constrained scenarios, is gradually gaining more and more attention, including but not limited to Neural Ar- chitecture Search [18, 19, 33, 35–40, 42, 43], network prun- ing [4, 7, 16, 27, 32, 41], and quantization [3, 5, 6, 15, 17, 21, 22,29,31]. Among these methods, quantization proposed to transform float network activations and weights to low-bit fixed points, which is capable of accelerating inference [13] or training [44] speed with little performance degradation. In general, network quantization methods are divided into quantization-aware training (QAT) [3, 5, 6] and post- training quantization (PTQ) [11, 17, 22, 31]. The former reduces the quantization error by quantization fine-tuning. Despite the remarkable results, the massive data require- 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. 7950 ments and high computational costs hinder the pervasive deployment of DNNs, especially on resource-constrained devices. Therefore, PTQ is proposed to solve the aforemen- tioned problem, which requires only minor or zero calibra- tion data for model reconstruction. Since there is no iter- ative process of quantization training, PTQ algorithms are extremely efficient, usually obtaining and deploying quan- tized models in a few minutes. However, this efficiency of- ten comes at the partial sacrifice of accuracy. PTQ typically performs worse than full precision models without quanti- zation training, especially in low-bit compact model quan- tization. Some recent algorithms [17, 22, 31] try to address this problem. For example, Nagel et al. [22] constructs new optimization functions by second-order Taylor expansions of the loss functions before and after quantization, which introduces soft quantization with learnable parameters to achieve adaptive weight rounding. Li et al. [17] changes layer-by-layer to block-by-block reconstruction and uses di- agonal Fisher matrices to approximate the Hessian matrix to retain more information. Wei et al. [31] discovers that ran- domly disabling some elements of the activation quantiza- tion can smooth the loss surface of the quantization weights. However, we observe that all the above methods show different degrees of oscillation with the deepening of the layer or block during the reconstruction process, as illus- trated in the left sub-figure of Fig. 1. We argue that the problem is essential and has been overlooked in the previ- ous PTQ methods. In this paper, through strict mathemati- cal definitions and proofs, we answer 3questions about the oscillation problem, which are listed as follows: (i).Why the oscillation happens in PTQ? To answer this question, we first define module topological homogene- ity, which relaxes the module equivalence restriction to a certain extent. And then, we give the definition of module capacity under the condition of module topological homo- geneity. In this case, we can prove that when the capacity of the later module is large enough, the reconstruction loss will break through the effect of quantization error accumulation and decrease. On the contrary, if the capacity of the later module is smaller than that of the preceding module, the reconstruction loss increases sharply due to the amplified quantization error accumulation effect. Overall, we demon- strate that the oscillation of the loss during PTQ reconstruc- tion is caused by the difference in module capacity; (ii).How the oscillation will influence the final perfor- mance? We observe that the final reconstruction error is highly correlated with the largest reconstruction error in all the previous modules by randomly sampling a large num- ber of mixed reconstruction granularity schemes. In other words, when oscillation occurs, the previous modules ob- viously have larger reconstruction errors, thus leading to worse accuracy in PTQ; (iii). How to solve the oscillation problem in PTQ?Since oscillation is caused by the different capacities of the front and rear modules, we propose the Mixed REC onstruction Granularity (MRECG) method which jointly optimizes the modules where oscillation occurs. Besides, our method is applicable in data-free and data- dependent scenarios, which is also compatible with differ- ent PTQ methods. In general, our contributions are listed as follows: • We reveal for the first time the oscillation problem in PTQ, which has been neglected in previous algorithms. However, we discover that smoothing out this oscilla- tion is essential in the optimization of PTQ. • We show theoretically that this oscillation is caused by the difference in the capability of adjacent modules. A small module capability exacerbates the cumulative effect of quantization errors making the loss increase rapidly, while a large module capability reduces the cu- mulative quantization errors making the loss decrease. • To solve the oscillation problem, we propose a novel Mixed REC onstruction Granularity (MRECG) method, which employs loss metric and module capac- ity to optimize mixed reconstruction granularity under data-dependency and data-free scenarios. The former finds the global optimum with moderately higher over- head and thus has the best performance. The latter is more effective with a minor performance drop. • We validate the effectiveness of the proposed method on a wide range of compression tasks in ImageNet. In particular, we achieve a Top- 1accuracy of 58.49% in MobileNetV 2with2/4bit, which exceeds current SOTA methods by a large margin. Besides, we also confirm that our algorithm indeed eliminates the oscil- lation of reconstruction loss on different models and makes the reconstruction process more stable.

Ofri-Amar_Neural_Congealing_Aligning_Images_to_a_Joint_Semantic_Atlas_CVPR_2023

Abstract We present Neural Congealing – a zero-shot self- supervised framework for detecting and jointly aligning semantically-common content across a given set of images. Our approach harnesses the power of pre-trained DINO- ViT features to learn: (i) a joint semantic atlas – a 2D grid that captures the mode of DINO-ViT features in the input set, and (ii) dense mappings from the unified atlas to each of the input images. We derive a new robust self- supervised framework that optimizes the atlas representa- tion and mappings per image set, requiring only a few real- world images as input without any additional input infor- mation (e.g., segmentation masks). Notably, we design our losses and training paradigm to account only for the shared content under severe variations in appearance, pose, back- ground clutter or other distracting objects. We demon- strate results on a plethora of challenging image sets in- cluding sets of mixed domains (e.g., aligning images depict- ing sculpture and artwork of cats), sets depicting related yet different object categories (e.g., dogs and tigers), or do- mains for which large-scale training data is scarce (e.g., coffee mugs). We thoroughly evaluate our method and show that our test-time optimization approach performs favor- ably compared to a state-of-the-art method that requires ex- tensive training on large-scale datasets. Project webpage: https://neural-congealing.github.io/

1. Introduction Humans can easily associate and match semantically- related objects across images, even under severe variations in appearance, pose and background content. For exam- ple, by observing the images in Fig. 1, we can immediately focus and visually compare the different butterflies, while ignoring the rest of the irrelevant content. While compu- tational methods for establishing semantic correspondences have seen a significant progress in recent years, research ef- forts are largely focused on either estimating sparse match- ing across multiple images (e.g., keypoint detection), or es- tablishing dense correspondences between a pair of images . In this paper, we consider the task of joint dense semantic alignment of multiple images . Solving this long-standing task is useful for a variety of applications, ranging from editing image collections [36,56], browsing images through canonical primitives, and 3D reconstruction (e.g., [8, 47]). The task of joint image alignment dates back to the sem- inal congealing [13, 14, 20, 28], which aligns a set of im- ages into a common 2D space. Recently, GANgealing [36] has modernized this approach for congealing an entire do- main of images. This is achieved by leveraging a pre-trained GAN to generate images that serve as self-supervisory sig- nal. Specifically, their method jointly learns both the mode of the generated images in the latent space of the GAN, and a network that predicts the mappings of the images into 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. 19403 joint mode. GANgealing demonstrated impressive results on in-the-wild image sets. Nevertheless, their method re- quires a StyleGAN model pre-trained on the domain of the test images, e.g., aligning cat images requires training Style- GAN on a large-scale cat dataset. This is a challenging task by itself, especially for unstructured image domains or un- curated datasets [34]. Moreover, they require additional ex- tensive training for learning the mode and their mapping network (e.g., training on millions of generated images). In this work, we take a different route and tackle the joint alignment task in the challenging setting where only a test image set is available, without any additional training data. More specifically, given only a few images as input (e.g., <25 images), our method estimates the mode of the test set and their joint dense alignment, in a self-supervised manner. We assume the input images share a common semantic content, yet may depict various factors of variations, such as pose, appearance, background content or other distracting objects (e.g., Mugs in Fig. 3). We take inspiration from the tremendous progress in representation learning, and lever- age a pre-trained DINO-ViT – a Vision Transformer model trained in a self-supervised manner [4]. DINO-ViT features have been shown to serve as an effective visual descriptor, capturing localized and semantic information (e.g., [2,49]). Here, we propose a new self-supervised framework that jointly and densely aligns the images in DINO-ViT feature space . To the best of our knowledge, we are the first to har- ness the power of DINO-ViT for dense correspondences be- tween in-the-wild images. More specifically, given an im- age set, our framework estimates, at test-time: (i) a joint la- tent 2D atlas that represents the mode of DINO-ViT features across the images, and (ii) dense mappings from the atlas to each of the images. Our training objective is driven by a matching loss encouraging each image features to match the canonical learned features in the joint atlas. We further incorporate additional loss terms that allow our framework to robustly represent and align only the shared content in the presence of background clutter or other distracting objects. Since our atlas and mappings are optimized per set, our method works in a zero-shot manner and can be applied to a plethora of image sets, including sets of mixed do- mains (e.g., aligning images depicting sculpture and art- work of cats), sets depicting related yet different object categories (e.g., dogs and tigers), or domains for which a dedicated generator is not available (e.g., coffee mugs). We thoroughly evaluate our method, and demonstrate that our test-time optimization framework performs favorably compared to [36] and on-par with state-of-the-art self- supervised methods. We further demonstrate how our atlas and mappings can be used for editing the image set with minimal effort by automatically propagating edits that are applied to a single image to the entire image set.

Luo_Constrained_Evolutionary_Diffusion_Filter_for_Monocular_Endoscope_Tracking_CVPR_2023

Abstract Stochastic filtering is widely used to deal with nonlin- ear optimization problems such as 3-D and visual track- ing in various computer vision and augmented reality ap- plications. Many current methods suffer from an imbal- ance between exploration and exploitation due to their par- ticle degeneracy and impoverishment, resulting in local op- timums. To address this imbalance, this work proposes a new constrained evolutionary diffusion filter for nonlinear optimization. Specifically, this filter develops spatial state constraints and adaptive history-recall differential evolu- tion embedded evolutionary stochastic diffusion instead of sequential resampling to resolve the degeneracy and im- poverishment problem. With application to monocular en- doscope 3-D tracking, the experimental results show that the proposed filtering significantly improves the balance be- tween exploration and exploitation and certainly works bet- ter than recent 3-D tracking methods. Particularly, the sur- gical tracking error was reduced from 4.03 mm to 2.59 mm.

1. Introduction Tracking a camera’s 3-D motion is vital in various com- puter vision applications, e.g., augmented reality, 3-D re- construction, computer assisted surgery, navigation and mapping, and robotics. Recent advances in 3-D tracking are widely discussed in the literature [9, 11, 13, 20, 21, 33]. Different from commonly used cameras in daily life, en- doscopic cameras are typical hand-held devices (called en- doscopes) used to inspect interior surfaces or inaccessible regions of tubular or hollow structures where the human visual system can hardly observe. While industrial endo- scopes are powerful for examining unreachable areas of *The author would like to give his special thanks to Professor Raymond Honfu Chan who is with Hong Kong Centre for Cerebro-cardiovascular Health Engineering and City University of Hong Kong. This work was supported in part by the National Nature Science Foundation of China un- der Grants 82272133 and 61971367, in part by the Fujian Provincial Tech- nology Innovation Joint Funds under Grant 2019Y9091, and in part by the Fujian Provincial Natural Science Foundation under Grant 2020J01004.buildings or parts of machines, surgical endoscopes are use- ful to intuitively inspect cavities in the body. Monocular endoscopic 3-D tracking plays an essential role in precise industrial inspection, clinical diagnosis and treatment. Unfortunately, surgical endoscopic cameras only provide 2-D video images without any depth information and cannot localize themselves and targets of interest like tumors in the surgical field. To this end, surgical 3-D tracking methods are widely developed to accurately localize surgical tools and targets and reduce inadvertent hurts in endoscopic or robotic surgery [16, 19, 26]. Such 3-D tracking is a nonlin- ear optimization problem as well as a multisensor or mul- tisource information fusion procedure, which is commonly solved by stochastic optimization methods [4]. Stochastic filtering is widely used for 3-D tracking [23], and usually generates a population of particles (initial solu- tions) and propagates them to approximate the optimal solu- tion. But it still limits itself to local optimums or premature convergence due to an imbalance between exploration and exploitation. Specifically, this imbalance results from the particle degeneracy and impoverishment after sequential re- sampling, leading to ineffective filtering. Theoretically, this work aims to solve the particle degeneracy-impoverishment problem to balance exploring and exploiting and create a new effective and powerful filtering strategy with robust op- timization performance. Technically, this work also strives for addressing several challenges in current surgical 3-D tracking methods: (1) endoscopic image uncertainty or arti- facts in vision-based 3-D tracking, (2) inaccurate and jitter measurements in sensor-based 3-D tracking, and (3) tissue deformation and patient movement in surgical procedures. Technical contributions of this work are clarified as fol- lows. First of all, two new spatial state constraints are in- troduced for nonlinear optimization problems, improving the optimization performance. More interestingly, a new strategy of evolutionary stochastic diffusion with adaptive history-recall differential evolution instead of sequential re- sampling can successfully resolve the particle degeneracy- impoverishment problem, effectively balancing between ex- ploration and exploitation. We then propose constrained 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. 4747 evolutionary diffusion filtering (CEDF), which is a meta- heuristic optimization algorithm and more ambidextrous than other filters. Additionally, a new hybrid bronchoscope 3-D tracking framework using the proposed filtering is de- veloped to fuse multisource data including computed to- mography (CT) or magnetic resonance (MR) images, surgi- cal videos, and positional sensor measurements. Our frame- work can tackle these challenges discussed above.

Li_Source-Free_Video_Domain_Adaptation_With_Spatial-Temporal-Historical_Consistency_Learning_CVPR_2023

Abstract Source-free domain adaptation (SFDA) is an emerging research topic that studies how to adapt a pretrained source model using unlabeled target data. It is derived from unsu- pervised domain adaptation but has the advantage of not requiring labeled source data to learn adaptive models. This makes it particularly useful in real-world applications where access to source data is restricted. While there has been some SFDA work for images, little attention has been paid to videos. Naively extending image-based methods to videos without considering the unique properties of videos often leads to unsatisfactory results. In this paper, we pro- pose a simple and highly flexible method for Source-Free Video Domain Adaptation (SFVDA), which extensively ex- ploits consistency learning for videos from spatial, tempo- ral, and historical perspectives. Our method is based on the assumption that videos of the same action category are drawn from the same low-dimensional space, regardless of the spatio-temporal variations in the high-dimensional space that cause domain shifts. To overcome domain shifts, we simulate spatio-temporal variations by applying spatial and temporal augmentations on target videos and encour- age the model to make consistent predictions from a video and its augmented versions. Due to the simple design, our method can be applied to various SFVDA settings, and ex- periments show that our method achieves state-of-the-art performance for all the settings.

1. Introduction Action recognition is a crucial task in video understand- ing and has been receiving tremendous attention from the vision community. In recent years, it has made significant progress, primarily due to the development of deep learning techniques [11, 43,45] and the establishment of large-scale annotated datasets [2, 13,42]. However, it is acknowledged that an action recognition model trained with annotated data drawn from one distribution typically experiences a per-formance drop when tested on out-of-distribution data [4]. This is the so-called domain shift problem. To tackle this problem, Unsupervised Video Domain Adaptation (UVDA) has been proposed. The goal is to learn an adaptive model using labeled video data from one domain (source) and unlabeled video data from another do- main (target). Typical UVDA methods use videos from both domains as input and train a model by minimizing the clas- sification risk on labeled source videos and explicitly align- ing videos from both domains in a class-agnostic fashion. Although most image-based domain alignment techniques can be applied to video domain alignment, such as adver- sarial learning [22, 37,44], methods that align domains by considering the richer temporal information in videos have shown superior performance [6, 33,36]. While UVDA methods help alleviate the domain shift problem, their assumption that labeled source videos are available for domain alignment can be problematic in real- world applications where access to source videos is re- stricted due to privacy or commercial reasons [24, 50]. This motivates a new research topic, Source-Free Video Domain Adaptation (SFVDA) [50], which aims to learn an adap- tive action recognition model using unlabeled target videos and a source model pre-trained with labeled source videos. SFVDA is similar to UVDA in learning an adaptive model using labeled source and unlabeled target videos but dif- fers in that labeled source videos are only used for learning the source model. Adaptation only involves target videos, which avoids leaking annotated source videos. However, the absence of labeled source videos makes SFVDA a more challenging problem than UVDA since there is no reliable supervision signal, and no data drawn from the distribution to be aligned, which makes it even more challenging. Very recently, Xu et al. [50] proposed a pioneering ap- proach to SFVDA based on temporal consistency. They adapt the source model by encouraging it to keep the ca- pability of understanding motion dynamics despite domain shifts. They train the model to produce features/predictions for a video clip consistent with those of other clips within the same video or that of the entire video. Despite im- 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. 14643 Spatial SpatialTemporalTemporalFigure 1. Conceptual illustration of applying spatial and temporal augmentations to simulate domain shifts and encouraging predic- tion consistency for SFVDA. proved performance over baseline methods, this method only considers adapting the source model from a tempo- ral perspective and ignores spatial factors (the appearance of frames) that also account for domain shifts. Adapting the model without encouraging it to surpass the visual ap- pearance variations could still produce sub-optimal adapta- tion results. Besides, clips from the same video often share high similarity, and the model can produce consistent fea- tures/predictions even though it has not been well adapted. In this paper, we propose a novel SFVDA method that overcomes the limitation of the existing methods by ex- ploiting Spatial-Temporal-Historical Consistency (STHC). Our underlying assumption is that videos of the same ac- tion category are drawn from the same low-dimensional space, regardless of spatio-temporal variations in the high- dimensional space that cause domain shifts . To achieve this, we simulate spatio-temporal variations with target videos and adapt the source model by encouraging it to surpass the variations and produce consistent predictions. Specifically, we apply spatial and temporal augmentations to each unlabeled target video in a stochastic manner to sim- ulate spatio-temporal variations. By encouraging consistent classification predictions for the video and its augmented versions, we ensure that they are drawn from the same low- dimensional space. After adapting the model in this way, it is expected to generalize well on test videos that fall into the same low-dimensional space as the training videos. Figure 1 provides an illustration of this concept. More concretely, we randomly select a clip from the video and apply stochastic frame-wise spatial augmenta- tion, resulting in a perturbed version of the clip. In addi- tion, we also apply stochastic temporal augmentation by randomly masking some frames to generate a temporally- perturbed clip. To ensure prediction consistency, we en- force the spatial consistency (SC) of the clip with its per- turbed version and the temporal consistency (TC) of the clip with its temporally-perturbed version. Besides these two techniques, we propose a third technique that enforces consistent predictions for the clip and other clips from the same video. This technique is similar to that in [50], butwe implement this in a nearly no-cost way: We store his- torical predictions of all the clips (with randomly sampled frames) from each video in a memory bank and retrieve pre- dictions from the bank to enforce prediction consistency for the current clip. This technique reinforces temporal consis- tency and we call it historical consistency (HC) . Notably, TC and SC produce “hard” versions of a clip and encourage the model to overcome the hard factors and make consis- tent predictions. Therefore, the model must have a strong understanding of the target domain to fulfill these tasks, fa- cilitating model adaptation. Thanks to simplicity in design, our STHC method can be easily extended to other SFVDA settings, including the open-set setting where the target domain contains classes that are absent in the source domain, the partial setting where the source domain contains classes that are absent in the target domain, and the black-box setting where only outputs of the source model are available and the model weights are not accessible. Experiments show that STHC outperforms existing methods for all the SFVDA settings. Our contributions can be summarized as follows: • We comprehensively exploit consistency learning for videos and propose STHC model for SFVDA. STHC performs stochastic spatio-temporal augmentations on each video and enforces prediction consistency from spatial, temporal, and historical perspectives. • We extend STHC to address various domain adap- tation problems under the SFVDA setting. To our best knowledge, most of these problems have not been studied before and we establish the evaluation bench- marks that will help future development. • STHC achieves state-of-the-art performance for SFVDA in various problem settings.

Mei_Unsupervised_Deep_Probabilistic_Approach_for_Partial_Point_Cloud_Registration_CVPR_2023

Abstract Deep point cloud registration methods face challenges to partial overlaps and rely on labeled data. To address these issues, we propose UDPReg, an unsupervised deep probabilistic registration framework for point clouds with partial overlaps. Specifically, we first adopt a network to learn posterior probability distributions of Gaussian mix- ture models (GMMs) from point clouds. To handle partial point cloud registration, we apply the Sinkhorn algorithm to predict the distribution-level correspondences under the constraint of the mixing weights of GMMs. To enable unsu- pervised learning, we design three distribution consistency- based losses: self-consistency, cross-consistency, and local contrastive. The self-consistency loss is formulated by en- couraging GMMs in Euclidean and feature spaces to share identical posterior distributions. The cross-consistency loss derives from the fact that the points of two partially over- lapping point clouds belonging to the same clusters share the cluster centroids. The cross-consistency loss allows the network to flexibly learn a transformation-invariant pos- terior distribution of two aligned point clouds. The lo- cal contrastive loss facilitates the network to extract dis- criminative local features. Our UDPReg achieves competi- tive performance on the 3DMatch/3DLoMatch and Model- Net/ModelLoNet benchmarks.

1. Introduction Rigid point cloud registration aims at determining the optimal transformation to align two partially overlapping point clouds into one coherent coordinate system [21, 30– 32]. This task dominates the performance of systems in many areas, such as robotics [57], augmented reality [6], autonomous driving [35, 42], radiotherapy [27], etc. Re- cent advances have been monopolized by learning-based approaches due to the development of 3D point cloud rep- resentation learning and differentiable optimization [37]. Existing deep learning-based point cloud registration methods can be broadly categorized as correspondence-free[2, 21, 30, 32, 47] and correspondence-based [4, 9, 19, 50]. The former minimizes the difference between global features extracted from two input point clouds. These global features are typically computed based on all the points of a point cloud, making correspondence-free ap- proaches inadequate to handle real scenes with partial over- lap [9, 55]. Correspondence-based methods first extract lo- cal features used for the establishment of point-level [9, 17, 19,21] or distribution-level [15,29,39,52] correspondences, and finally, estimate the pose from those correspondences. However, point-level registration does not work well un- der conditions involving varying point densities or repeti- tive patterns [31]. This issue is especially prominent in in- door environments, where low-texture regions or repetitive patterns sometimes dominate the field of view. Distribution- level registration, which compensates for the shortcomings of point-level methods, aligns two point clouds without es- tablishing explicit point correspondences. Unfortunately, to the best of our knowledge, the existing methods are inflex- ible and cannot handle point clouds with partial overlaps in real scenes [28, 31]. Moreover, the success of learning- based methods mainly depends on large amounts of ground truth transformations or correspondences as the supervision signal for model training. Needless to say, the required ground truth is typically difficult or costly to acquire, thus hampering their application in the real world [38]. We thus propose an unsupervised deep probabilistic reg- istration framework to alleviate these limitations. Specif- ically, we extend the distribution-to-distribution (D2D) method to solve partial point cloud registration by adopt- ing the Sinkhorn algorithm [11] to predict correspondences of distribution. In order to make the network learn ge- ometrically and semantically consistent features, we de- sign distribution-consistency losses, i.e., self-consistency and cross-consistency losses, to train the networks without using any ground-truth pose or correspondences. Besides, we also introduce a local contrastive loss to learn more dis- criminative features by pushing features of points belong- ing to the same clusters together while pulling dissimilar features of points coming from different clusters apart. 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. 13611 Our UDPReg is motivated by OGMM [33] and UGMM [20] but differs from them in several ways. Firstly, unlike OGMM, which is a supervised method, our approach is unsupervised. Secondly, while UGMM [20] treats all clusters equally in the matching process, our method aligns different clusters with varying levels of importance. This enables our approach to handle partial point cloud regis- tration successfully. To enable unsupervised learning, the designed self-consistency loss encourages the extracted fea- tures to be geometrically consistent by compelling the fea- tures and coordinates to share the posterior probability. The cross-consistency loss prompts the extracted features to be geometrically consistent by forcing the partially overlap- ping point clouds to share the same clusters. We evaluate our UDPReg on 3DMatch [53], 3DLoMatch [19], Model- Net [45] and ModelLoNet [19], comparing our approach against traditional and deep learning-based point cloud reg- istration approaches. UDPReg achieves state-of-the-art re- sults and significantly outperforms unsupervised methods on all the benchmarks. In summary, the main contributions of this work are: • We propose an unsupervised learning-based probabilistic framework to register point clouds with partial overlaps. • We provide a deep probabilistic framework to solve par- tial point cloud registration by adopting the Sinkhorn al- gorithm to predict distribution-level correspondences. • We formulate self-consistency, cross-consistency, and local-contrastive losses, to make the posterior probabil- ity in coordinate and feature spaces consistent so that the feature extractor can be trained in an unsupervised way. • We achieve state-of-the-art performance on a compre- hensive set of experiments, including synthetic and real- world datasets1.

Li_MoDAR_Using_Motion_Forecasting_for_3D_Object_Detection_in_Point_CVPR_2023

Abstract Occluded and long-range objects are ubiquitous and challenging for 3D object detection. Point cloud sequence data provide unique opportunities to improve such cases, as an occluded or distant object can be observed from differ- ent viewpoints or gets better visibility over time. However, the efficiency and effectiveness in encoding long-term se- quence data can still be improved. In this work, we propose MoDAR, using motion forecasting outputs as a type of vir- tual modality, to augment LiDAR point clouds. The MoDAR modality propagates object information from temporal con- texts to a target frame, represented as a set of virtual points, one for each object from a waypoint on a forecasted tra- jectory. A fused point cloud of both raw sensor points and the virtual points can then be fed to any off-the-shelf point- cloud based 3D object detector. Evaluated on the Waymo Open Dataset, our method significantly improves prior art detectors by using motion forecasting from extra-long se- quences (e.g. 18 seconds), achieving new state of the arts, while not adding much computation overhead.

1. Introduction 3D object detection is a fundamental task for many appli- cations such as autonomous driving. While there has been tremendous progress in architecture design and LiDAR- camera sensor fusion, occluded and long-range object de- tection remains a challenge. Point cloud sequence data pro- vide unique opportunities to improve such cases. In a dy- namic scene, as the ego-agent and other objects move, the sequence data can capture different viewpoints of objects or improve their visibility over time. The key challenge though, is how to efficiently and effectively leverage se- quence data for 3D object detection. Existing multi-frame 3D object detection methods often fuse sequence data at two different levels. At scene level, the most straightforward approach is to transform point clouds of different frames to a target frame using known equal contributions OOMsaturateFigure 1. 3D detection model performance vs.number of input frames. Naively adding more frames to existing methods, such as CenterPoint [59] and SWFormer [40], quickly plateaus the gains while our method, MoDAR, scales up to many more frames and gets much larger gains. L2 3D mAPH is computed by averaging vehicle and pedestrian L2 3D APH. ego motion poses [3, 40, 55, 59]. Each point can be dec- orated with an extra time channel to indicate which frame it is from. However, according to previous studies [7, 33] and our experiments shown in Fig. 1, it is difficult to fur- ther improve the detection model by including more input frames due to its large computation overhead as well as in- effective temporal data fusion at scene level (especially for moving objects). On the other side, 3D Auto Labeling [33] and MPPNet [7] propose to aggregate longer temporal con- texts at object level, which is more tractable as there are much less points from objects than those from the entire scenes. However, they also fail to scale up temporal con- text aggregation to long sequences due to efficiency issues or alignment challenges. In our paper, we propose to use motion forecasting to propagate object information from the past (and the future) to a target frame. The output of the forecasting model can be considered another (virtual) sensor modality to the detector model. Inspired by the naming of the LiDAR sensor, we name this new modality MoDAR , Motion forecasting based Detection And Ranging (see Fig. 2 for an example). Traditionally 3D object detection is a pre-processing step for a motion forecasting model, where the detector boxes are either used as input (for past frames) or learning targets (for future frames). In contrast, we use motion forecasting outputs as input to LiDAR-MoDAR multi-modal 3D 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. 9329 An occluded object A long-range object Caption: Red & Blue points: MoDAR points from motion forecasting; Gery points: LiDAR points from raw sensor MoDAR points (red) from forward prediction MoDAR points (blue) from reverse prediction LiDAR points (gary) from LiDAR Figure 2. 3D object detection from MoDAR and LiDAR points. MoDAR points (red and blue) are predicted object centers with ex- tra features such as sizes, semantic classes and confidence scores. Compared to LiDAR-only detectors, a multi-modal detector taking both LiDAR (gray) and MoDAR points can accurately recognize occluded and long-range objects that have few observed points. detectors. There are two major benefits of using a MoDAR sensor for 3D object detection from sequence data. First, motion forecasting can easily transform object information across very distant frames (8 seconds or longer). Such prop- agation is especially robust to occlusions as the forecast- ing models do not assume successful tracking for trajectory forecasting. Second, considering forecasting output as an- other sensor data source for 3D detection, it is a lightweight sensor modality, making long-term sequence data process- ing possible without much computation overhead. Specifically, in MoDAR, we represent motion forecast- ing output at the target frame as a set of virtual points (named as MoDAR points), one for each object from a way- point on a forecasted trajectory. The predicted object loca- tion is the 3D coordinate of the virtual point, while addi- tional information (such as object type, size, predicted head- ing, and confidence score) is encoded into the virtual point features. Each virtual point is appended with a time channel to indicate the context frames it uses for the motion fore- casting. For a target frame, we can use forecasted outputs from multiple context frames easily through a union of cor- responding virtual points. In an offboard/offline detection setup, we can use both forward prediction and reverse pre- diction (use future frames as input to the forecasting model) to combine information from the past and the future. For de- tection, we fuse the raw sensor points (from LiDARs) and the virtual points (from forecasting), and feed them to any off-the-shelf point cloud based 3D detector. In experiments, we use a MultiPath++ [42] motion forecasting model trained on the Waymo Open Motion Dataset [9] to generate MoDAR points from past 9 seconds for online detection; and from past and future 18 seconds for offline detection. With minimum changes, we adapt Cen- terPoints [59] and SWFormer [40] detectors for LiDAR- MoDAR 3D object detection.1Evaluated on the Waymo 1Although we experiment with point-cloud based detectors, MoDAROpen Dataset [39], we show that adding MoDAR signifi- cantly improves detection quality, improving CenterPoints and SWFormer by 11:1and 8:5mAPH respectively; and it especially helps detection of long-range and occluded objects. Using MoDAR with a 3-frame SWFormer detec- tor, we have achieved state-of-the-art mAPH on the Waymo Open Dataset. We further provide extensive ablations and analysis experiments to validate our designs and show im- pacts of different MoDAR choices.

Lu_Visual_Language_Pretrained_Multiple_Instance_Zero-Shot_Transfer_for_Histopathology_Images_CVPR_2023

Abstract Contrastive visual language pretraining has emerged as a powerful method for either training new language-aware image encoders or augmenting existing pretrained models with zero-shot visual recognition capabilities. However, existing works typically train on large datasets of image- text pairs and have been designed to perform downstream tasks involving only small to medium sized-images, neither of which are applicable to the emerging field of computa- tional pathology where there are limited publicly available paired image-text datasets and each image can span up to 100,000 ×100,000 pixels. In this paper we present MI- Zero, a simple and intuitive framework for unleashing the zero-shot transfer capabilities of contrastively aligned im- age and text models on gigapixel histopathology whole slide images, enabling multiple downstream diagnostic tasks to be carried out by pretrained encoders without requiring any additional labels. MI-Zero reformulates zero-shot trans- fer under the framework of multiple instance learning to overcome the computational challenge of inference on ex- tremely large images. We used over 550k pathology re- ports and other available in-domain text corpora to pre- train our text encoder. By effectively leveraging strong pre- trained encoders, our best model pretrained on over 33k histopathology image-caption pairs achieves an average median zero-shot accuracy of 70.2% across three different real-world cancer subtyping tasks. Our code is available at: https://github.com/mahmoodlab/MI-Zero.

1. Introduction Weakly-supervised deep learning for computational pathology (CPATH) has rapidly become a standard ap- proach for modelling whole slide image (WSI) data [9, 30, 47,71,73]. To obtain “clinical grade” machine learning per- formance on par with human experts for a given clinical †These authors contributed equally to this work.task, many approaches adopt the following model develop- ment life cycle: 1) curate a large patient cohort ( N > 1000 samples) with diagnostic whole-slide images and clinical labels, 2) unravel and tokenize the WSI into a sequence of patch features, 3) use labels to train a slide classifier that learns to aggregate the patch features for making a predic- tion, and 4) transfer the slide classifier for downstream clin- ical deployment [9, 43, 91]. Successful examples of task-specific model development (e.g. training models from scratch for each task) include prostate cancer grading and lymph node metastasis detec- tion [5, 7–9, 50, 70]. However, this paradigm is intractable if one wishes to scale across the hundreds of tumor types across the dozens of different organ sites in the WHO clas- sification system1, with most tumor types under-represented in public datasets or having inadequate samples for model development [41, 92]. To partially address these limita- tions, self-supervised learning has been explored for learn- ing the patch representations within the WSI with the idea that certain local features, such as tumor cells, lymphocytes, and stroma, may be conserved and transferred across tis- sue types [10, 16, 39, 40, 44, 64, 77]. Though morphological features at the patch-level are captured in a task-agnostic fashion, developing the slide classifier still requires supervi- sion, which may not be possible for disease types with small sample sizes. To scale slide classification across the vast number of clinical tasks and possible findings in CPATH, an important shift needs to be made from task-specific to task-agnostic model development. Recent works [33,55] have demonstrated that large-scale pretraining using massive, web-sourced datasets of noisy image-text pairs can not only learn well-aligned represen- tation spaces between image and language, but also transfer the aligned latent space to perform downstream tasks such as image classification. Specifically for CLIP [55], after pretraining a vision encoder in a task-agnostic fashion, the vision encoder can be “prompted” with text from the label 1tumourclassification.iarc.who.int/ 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. 19764 space (referred to as “zero-shot transfer”, as no labeled ex- amples are used in the transfer protocol). Despite the vol- ume of zero-shot transfer applications developed for natural images [21, 33, 45, 49, 57, 80, 82] and certain medical imag- ing modalities ( e.g.radiology [29,60,68,78,90]), zero-shot transfer for pathology has not yet been studied2. We believe this is due to 1) the lack of large-scale, publicly available datasets of paired images and captions in the highly special- ized field of pathology, and 2) fundamental computational challenges associated with WSIs, as images can span up to 100,000×100,000pixels and do not routinely come with textual descriptions, bounding box annotations or even re- gion of interest labels. In this work, we overcome the above data and compu- tational challenges and develop the first zero-shot transfer framework for the classification of histopathology whole slide images. On the data end, we curated the largest known dataset of web-sourced image-caption pairs specifically for pathology. We propose “MI-Zero”, a simple and intuitive multiple instance learning-based [3, 30] method for utiliz- ing the zero-shot transfer capability of pretrained visual- language encoders for gigapixel-sized WSIs that are rou- tinely examined during clinical practice. We validate our approach on 3 different real-world cancer subtyping tasks, and perform multiple ablation experiments that explore im- age pretraining, text pretraining, pooling strategies, and sample size choices for enabling zero-shot transfer in MI- Zero.

Lou_All-in-Focus_Imaging_From_Event_Focal_Stack_CVPR_2023

Abstract Traditional focal stack methods require multiple shots to capture images focused at different distances of the same scene, which cannot be applied to dynamic scenes well. Generating a high-quality all-in-focus image from a single shot is challenging, due to the highly ill-posed nature of the single-image defocus and deblurring problem. In this pa- per, to restore an all-in-focus image, we propose the event focal stack which is defined as event streams captured dur- ing a continuous focal sweep. Given an RGB image focused at an arbitrary distance, we explore the high temporal reso- lution of event streams, from which we automatically select refocusing timestamps and reconstruct corresponding refo- cused images with events to form a focal stack. Guided by the neighbouring events around the selected timestamps, we can merge the focal stack with proper weights and restore a sharp all-in-focus image. Experimental results on both synthetic and real datasets show superior performance over state-of-the-art methods. †Contributed equally to this work as first authors ∗Corresponding author Project page: https://hylz-2019.github.io/EFS

1. Introduction The lens aperture of a camera controls the amount of incoming luminous flux. A larger aperture maintains the signal-to-noise ratio with shorter exposure time, which is useful for shooting high-speed scenes or capturing images in low-light conditions with less noise. However, large aper- ture settings also make the depth of field (DoF) shallow, which results in defocus blur. This is preferable in certain scenarios, such as in portrait photography a shallow DoF can be used to emphasize the subject. Yet, all-in-focus im- ages preserve information from all distances and are desired in more situations, e.g., microscopy imaging [25]. Besides, all-in-focus imaging also benefits various high-level vision tasks, e.g., object detection [29] and semantic segmenta- tion [10]. An all-in-focus image could be obtained by deblurring a defocused image, but the defocus kernel, determined by the aperture shape and depth of the scene, is usually spatially- varying and difficult to be estimated accurately [48]. Con- ventional two-stage methods [9, 13, 36] first estimate the pixel-wise or patch-wise defocus kernels with image pri- ors and then apply non-blind image deconvolution to each 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. 17366 pixel or patch. Recently, benefiting from the data-driven strategy, end-to-end deep learning methods [18, 31, 32, 38] outperform conventional two-stage restoration methods, by observing defocused and all-in-focus image pairs during training. Although they have demonstrated high potential in removing defocus blur, the deblurred results still cannot avoid ringing artifacts or remain blurry in high-frequency regions due to inaccurate defocus kernel estimation espe- cially for weakly textured and defocused regions (an exam- ple is shown in Figure 1 right (c)). To overcome the ill-posedness of estimating the defocus kernel from a single image, merging a focal stack, i.e., a sequence of images taken at different focus distances, can generate an all-in-focus image reliably [11, 40, 47]. How- ever, capturing a focal stack requires a static scene and mul- tiple exposures. Moreover, the selection of focus distances is a key factor in capturing the focal stack, which requires elaborate design. Neuromorphic event cameras [5, 35] are novel sen- sors that can detect brightness changes and trigger an event whenever its log variation exceeds a preset thresh- old. Thanks to their high temporal resolution featured with microsecond-level sensitivity, they can capture ap- proximately continuous signals for intensity variations of a scene, and support applications like generating high-speed videos from event streams [28, 41–43]. These characteris- tics motivate us to think about: Can we use “focal stacks” composed of event streams for all-in-focus imaging? In this paper, we propose event focal stack (EFS) for the first time. It is composed of event streams obtained from a continuous focal sweep with an event camera, which can be used to reconstruct an image focal stack (given an RGB im- age focused at an arbitrary distance) and predict the merging weights for all-in-focus image recovery, as shown in Fig- ure 1 left. EFS encodes scene texture information from con- tinuous different depths in temporal log-gradient domain, so we first select a refocusing timestamp for each patch of the scene, which corresponds to sharper edges and richer tex- ture information at that time. By fusing a defocused image and the EFS recorded between the defocused timestamp and refocusing timestamp, we generate a refocused image for each refocusing timestamp, forming an image focal stack. Guided by neighbouring events around refocusing times- tamps, we can predict the merging weight for each image needed for composing a focal stack, and finally restore an all-in-focus image (an example is shown in Figure 1 right (d)). Contributions of this paper are demonstrated by ex- ploring the following benefits of the proposed EFS: • reliable selection of refocusing timestamps by decod- ing continuous scene gradient changes from events; • consistent link between defocused (given) and refo- cused images (estimated) composing an image focalstack; and • robust guidance for merging weight prediction and all- in-focus reproduction with event triggered neighbour- ing the selected timestamps. We quantitatively and qualitatively evaluate our method on both synthetic and real datasets and demonstrate its su- perior quality in recovering all-in-focus images over state- of-the-art methods.

Nitzan_Domain_Expansion_of_Image_Generators_CVPR_2023

Abstract Can one inject new concepts into an already trained gen- erative model, while respecting its existing structure and knowledge? We propose a new task – domain expansion – to address this. Given a pretrained generator and novel (but related) domains, we expand the generator to jointly model all domains, old and new, harmoniously. First, we note the generator contains a meaningful, pretrained latent space. Is it possible to minimally perturb this hard-earned representation, while maximally representing the new do- mains? Interestingly, we find that the latent space offers un- used, “dormant” directions, which do not affect the output. This provides an opportunity: By “repurposing” these di- rections, we can represent new domains without perturbing the original representation. In fact, we find that pretrained generators have the capacity to add several – even hundreds – of new domains! Using our expansion method, one “ex- panded” model can supersede numerous domain-specific models, without expanding the model size. Additionally, a single expanded generator natively supports smooth transi- tions between domains, as well as composition of domains. Code and project page available here.

1. Introduction Recent domain adaptation techniques piggyback on the tremendous success of modern generative image models [3, 12, 32, 40], by adapting a pretrained generator so it can generate images from a new target domain. Oftentimes, the target domain is defined with respect to the source do- main [5,21,22], e.g., changing the “stylization” from a pho- torealistic image to a sketch. When such a relationship holds, domain adaptation typically seeks to preserve the fac- tors of variations learned in the source domain, and transfer them to the new one (e.g., making the human depicted in a sketch smile based on the prior from a face generator). With existing techniques, however, the adapted model loses the ability to generate images from the original domain. In this work, we introduce a novel task — domain ex- pansion . Unlike domain adaptation, we aim to augment the space of images a single model can generate, without over- riding its original behavior (see Fig. 1). Rather than view- ing similar image domains as disjoint data distributions, we treat them as different modes in a joint distribution. As a result, the domains share a semantic prior inherited from the original data domain. For example, the inherent factors of variation for photorealistic faces, such as pose and face shape, can equally apply to the domain of “zombies”. 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. 15933 DogCuteSiberian HuskySketch Expanded DomainsBoarHappyPop ArtSource Domain Domain CompositionSiberian Husky + Cute + SketchBoar + Happy +Pop ArtFigure 2. Example of a domain expansion result. Starting from dogs as the source domain, we expand a single generator to model new domains such as facial expressions, breeds of dogs and other animals, and artistic styles. Finally, as the representations are dis- entangled, the expanded generator is able to generalize and com- pose the different domains, although they were never seen jointly in training. To this end, we carefully structure the model train- ing process for expansion, respecting the original data do- main. It is well-known that modern generative models with low-dimensional latent spaces offer an intriguing, emer- gent property – through training, the latent spaces represent the factors of variation, in a linear and interpretable man- ner [3, 6, 10, 12, 28, 30, 39, 40]. We wish to extend this ad- vantageous behavior and represent the new domains along linear and disentangled directions. Interestingly, it was pre- viously shown that many latent directions have insignificant perceptible effect on generated images [6]. Taking advan- tage of this finding, we repurpose such directions to repre- sent the new domains. In practice, we start from an orthogonal decomposition of the latent space [36] and identify a set of low-magnitude directions that have no perceptible effect on the generated images, which we call dormant . To add a new domain, we select a dormant direction to repurpose. Its orthogo- nal subspace, which we call base subspace , is sufficient to represent the original domain [6]. We aim to repur- pose the dormant direction such that traversing it would now cause a transition between the original and the new domain. Specifically, the transition should be disentangled from the original domain’s factors of variation. To this end, we define a repurposed affine subspace by transporting the base subspace along the chosen dormant direction, as shown in Fig. 3. We capture the new domain by applying a domain adaptation method, transformed to operate only on latent codes sampled from the repurposed subspace. A regular- ization loss is applied on the base subspace to ensure that the original domain is preserved. The original domain’s factors of variation are implicitly preserved due to the sub- spaces being parallel and the latent space being disentan- gled. For multiple new domains, we simply repeat this pro- cedure across multiple dormant directions. We apply our method to the StyleGAN [13] architecture,with multiple datasets, and expand the generator with hun- dreds of new factors of variation. Crucially, we show our expanded model simultaneously generates high-quality im- ages from both original and new domains, comparable to specialized, domain-specific generators. Thus, a single ex- panded generator supersedes hundreds of adapted genera- tors, facilitating the deployment of generative models for real-world applications. We additionally demonstrate that the new domains are learned as global and disentangled fac- tors of variation, alongside existing ones. This enables fine- grained control over the generative process and paves the way to new applications and capabilities, e.g., compositing multiple domains (See Fig. 2). Finally, we conduct a de- tailed analysis of key aspects of our method, such as the ef- fect of the number of newly introduced domains, thus shed- ding light on our method and, in the process, on the nature of the latent space of generative models. To summarize, our contributions are as follows: • We introduce a new task – domain expansion of a pre- trained generative model. • We propose a novel latent space structure that is amenable to representing new knowledge in a disentangled manner, while maintaining existing knowledge intact. • We present a simple paradigm transforming domain adap- tation methods into domain expansion methods. • We demonstrate successful domain expansion to hun- dreds of new domains and illustrate its advantage over domain adaptation methods.

Park_Mask-Guided_Matting_in_the_Wild_CVPR_2023

Abstract Mask-guided matting has shown great practicality com- pared to traditional trimap-based methods. The mask- guided approach takes an easily-obtainable coarse mask as guidance and produces an accurate alpha matte. To ex- tend the success toward practical usage, we tackle mask- guided matting in the wild , which covers a wide range of categories in their complex context robustly. To this end, we propose a simple yet effective learning framework based on two core insights: 1) learning a generalized matting model that can better understand the given mask guidance and 2) leveraging weak supervision datasets (e.g., instance segmentation dataset) to alleviate the limited diversity and scale of existing matting datasets. Extensive experimen- tal results on multiple benchmarks, consisting of a newly proposed synthetic benchmark (Composition-Wild) and ex- isting natural datasets, demonstrate the superiority of the proposed method. Moreover, we provide appealing results on new practical applications (e.g., panoptic matting and mask-guided video matting), showing the great generality and potential of our model.

1. Introduction Image matting aims to predict the opacity of ob- jects, which enables precise separation from surround- ing backgrounds. Due to the ill-posed nature of the task, many works [7, 13, 21, 27, 30, 48] have improved matting performance by relying on the manual guidance of a trimap . However, pixel-level annotation of fore- ground/background/unknown is extremely burdensome, re- stricting its usage in many practical applications such as im- age/video editing and film production. Recently, many ef- ficient alternatives for user guidance have been proposed, including trimap-free [15, 32], additional background im- ages [22, 34], scribble [43], and the user clicks [45]. Among them, the mask-guided approach [50] shows a great trade-off between performance and intensity of user interaction. It utilizes a coarse mask as guidance, which is much easier to obtain either manually or from off-the-shelf Image and Guidance Ours MGMattingFigure 1. Qualitative Comparisons of MGMatting [50] and Ours in the wild. The mask guidance is overlaid on images with blue color. Best viewed zoomed in. segmentation models [2, 10]. With only the coarse spatial prior, the mask-guided matting model [50] shows compara- ble or even better performance than the trimap-based com- petitors [13,17,21,27,48] on synthetic Composition-1k [48] and a real-world human matting dataset [50]. However, de- spite the encouraging results, we see the previous state-of- the-art model [50] struggles to obtain desirable alpha matte in complex real-world scenes (see Fig. 1). With this observation, we tackle mask-guided matting in the wild . Specifically, we formulate unique setups and emerging challenges of the new task as follows: (1) We aim to handle objects in their complex context, reflecting the characteristics of natural images. The previous method [50] evaluates their model on iconic-object images [1,29] where only a single object is in the center. As the model can easily find the target object in such images, the model’s real instance discrimination ability is, in fact, veiled. On the contrary, in an ‘in-the-wild’ setting, it is crucial to pre- cisely localize the target object from the given coarse/noisy mask guidance ( i.e.mask awareness). (2) Our model tar- gets to deal with diverse categories of objects in natural 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. 1992 images. Unlike most previous methods that improve gen- eralization performance at the expense of category-specific regime ( e.g., limiting to humans [15] or animals [19]), we aim to understand distinctive matting patterns of vast cat- egories. (3) Limited data problem makes the new setting more complicated. Due to the labeling complexity, anno- tating alpha matte for objects in common scenes, e.g., the COCO dataset [24], is infeasible. As a sidestep, previ- ous benchmarks [32, 48] extract the alpha matte and fore- ground colors from images with simple backgrounds. These are composited on various backgrounds [6, 24], and result- ing samples are used to train and evaluate matting models. However, due to the inevitable composition artifacts, the models usually show limited generalization performance. In that sense, how to train and evaluate the in-the-wild mat- ting model remains an open question. Toward this goal, we propose a simple yet effective learning framework for a generalized mask-guided matting model. First, we investigate fundamental reasons for the poor generalization of the previous mask-guided matting model [50] and find that this is mainly from the training data generation process. Specifically, the previous compo- sition process includes instance merging data augmentation, which merges several foreground objects into a single ob- ject. While this augmentation is effective in the trimap- based methods [21, 30, 41], it implicitly makes a negative bias for the mask-guided matting model to ignore the guid- ance. Thus, the model struggles to localize the target objects in complex natural scenes. We alleviate the bias by propos- ing an instance-wise learning objective, where the model is supervised to segment one of multiple instances according to the guidance. By doing so, the model learns strong se- mantic representation regarding complex relations and soft transitions between objects. Despite the simplicity of the proposal, this greatly improves performance in the wild. Second, we explore a practical solution to make the mask-guided model handle various categories of objects ro- bustly. Instead of scaling the matting dataset, we leverage a dataset with weak supervision [14, 46] ( i.e., instance seg- mentation dataset [24]), as the coarse instance masks are easier to obtain over the diverse categories of objects. To effectively hallucinate the fine supervision signal with the weak localization guidance, we come up with a self-training framework [36, 37, 47]. Specifically, a pseudo label is gen- erated based on a weakly-augmented input (both image and instance mask annotation as guidance), which supervises the model prediction on a strongly augmented version of them. During self-training, the model is not only adapting to the in-the-wild scenario in a self-evolving manner but also being robust to noise in both image and guidance. To verify the in-the-wild performance of mask-guided matting, we formally define an evaluation protocol involv- ing multiple sub-benchmarks: Composition-Wild, AIM-500 [20], COCO [24]. We first design an in-the-wild ex- tension of the popular synthetic Composition-1k bench- mark [48], namely Composition-Wild. We simulate com- plex real-world images by compositing multiple foreground objects. To bring valuable insight on the failure cases of the model, we design sub-metrics for Composition-Wild. In addition, we use the AIM-500 dataset to establish quan- titative results on natural images ( i.e., with no composition artifacts), although most images are iconic-object images with simple backgrounds. Finally, we provide qualitative outputs of our mask-guided matting model on the COCO dataset [24] which is one of the most representative in-the- wild datasets. To summarize, we make the following main contribu- tions. 1) To our best knowledge, it is the first work to explore mask-guided matting in the wild. 2) We develop a simple yet effective learning framework leveraging both composited and weak-guidance images. 3) We design an evaluation setup for the new task. 4) We initiate several in- teresting extensions: video and panoptic matting.

Li_Open-Set_Semantic_Segmentation_for_Point_Clouds_via_Adversarial_Prototype_Framework_CVPR_2023

Abstract Recently, point cloud semantic segmentation has attracted much attention in computer vision. Most of the existing works in literature assume that the training and testing point clouds have the same object classes, but they are gen- erally invalid in many real-world scenarios for identifying the 3D objects whose classes are not seen in the training set. To address this problem, we propose an Adversarial Prototype Framework (APF) for handling the open-set 3D semantic segmentation task, which aims to identify 3D unseen-class points while maintaining the segmentation performance on seen-class points. The proposed APF consists of a feature extraction module for extracting point features, a prototypical constraint module, and a feature adversarial module. The prototypical constraint module is designed to learn prototypes for each seen class from point features. The feature adversarial module utilizes generative adversarial networks to estimate the distribution of unseen- class features implicitly, and the synthetic unseen-class features are utilized to prompt the model to learn more effective point features and prototypes for discriminating unseen-class samples from the seen-class ones. Experi- mental results on two public datasets demonstrate that the proposed APF outperforms the comparative methods by a large margin in most cases.

1. Introduction Point cloud semantic segmentation is an important and challenging topic in computer vision. Most of the existing works [9–11, 29] in literature are based on the assumption that both the training and testing point clouds have the same *Corresponding author (a) AD (b) O3D Figure 1. Visualization of the goals of anmaly detection (AD) and open-set 3D semantic segmentation (O3D) on SemanticKITTI [2]. AD is to identify the unseen-class data, while O3D is to simultane- ously identify the unseen-class data and segment seen-class data. The unseen-class points are colorized in yellow. object classes, however, this assumption is no more valid in many real-world scenarios, due to the fact that the classes of some observed 3D points may not be presented in the training set. Hence, the following problem on open-set 3D semantic segmentation is naturally raised: How does a seg- mentation model simultaneously identify unseen-class 3D points and maintain the segmentation accuracy of seen-class 3D points in open-set scenarios? Compared with anomaly detection [3, 23, 26], open-set 3D semantic segmentation (O3D) is more challenging, for it also needs to assign labels to seen-class data simultane- ously, as shown in Figure 1. In fact, some existing tech- niques [6,15,17,18] for open-set 2D semantic segmentation (O2D) task could be extended to handle the O3D task, how- ever, their open-set ability is generally limited in 3D scenar- ios. In addition, to our best knowledge, only one pioneering work [7] has investigated a special technique for O3D task. In [7], an O3D method called REAL is proposed to utilize normal classifiers to segment seen-class points and regard the randomly resized objects as unseen-class objects which are detected by the redundancy classifiers. REAL outper- forms some extended O2D methods in the O3D task, how- 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. 9425 ever, the AUPR (Area Under the Precision-Recall curve) is lower than 21% on two public datasets as shown in Ta- ble 1 and Table 2 in Section 4, mainly because the resizing process in REAL alters the geometric structure of the ini- tial point clouds to some extent. These results indicate that there still exists a huge space for improvement on O3D task. Addressing the above issue, we propose an Adversarial Prototype Framework (APF) for open-set 3D segmentation, which segments point clouds from a discriminative perspec- tive and estimates the distribution of unseen-class features from a generative perspective. The proposed APF consists of three modules: a feature extraction module, a prototyp- ical constraint module, and a feature adversarial module. The feature extraction module is employed to extract latent features from the input point clouds, which could be an ar- bitrary closed-set point cloud segmentation network in prin- ciple. Given the point features, the prototypical constraint module is explored from the discriminative perspective to learn a prototype for each seen class. The feature adver- sarial module is explored from the generative perspective, which employs the generative adversarial networks (GAN) to synthesize point features to estimate the unseen-class feature distribution, based on the finding stated in [6] that the unseen-class features usually aggregate in the center of the feature space. And the synthesized unseen-class fea- tures in this module could further prompt the model to learn more discriminative point features and prototypes. After the whole APF is trained, a point-to-prototype hybrid distance- based criterion is introduced for open-set 3D segmentation. In sum, the contributions of this paper are as follows: • We propose the adversarial prototype framework (APF) for handling the open-set 3D semantic segmen- tation task. Under the proposed APF, various open-set 3D segmentation methods could be straightforwardly derived by utilizing existing closed-set 3D segmenta- tion networks as the feature extraction module. The effectiveness of the proposed APF has been demon- strated by the experimental results in Section 4. • Under the proposed framework, we explore the pro- totypical constraint module, which learns the corre- sponding prototype for each seen class. The learned prototypes are not only conducive to segmenting seen- class points, but also to detecting unseen-class points. • Under the proposed framework, we explore the fea- ture adversarial module to synthesize unseen-class fea- tures. The synthetic features are helpful for improving the discriminability of both the seen-class features and prototypes via the adversarial mechanism.

Li_Robust_Model-Based_Face_Reconstruction_Through_Weakly-Supervised_Outlier_Segmentation_CVPR_2023

Abstract In this work, we aim to enhance model-based face re- construction by avoiding fitting the model to outliers, i.e. regions that cannot be well-expressed by the model such as occluders or make-up. The core challenge for localizing outliers is that they are highly variable and difficult to anno- tate. To overcome this challenging problem, we introduce a joint Face-autoencoder and outlier segmentation approach (FOCUS). In particular, we exploit the fact that the outliers cannot be fitted well by the face model and hence can be localized well given a high-quality model fitting. The main challenge is that the model fitting and the outlier segmenta- tion are mutually dependent on each other, and need to be inferred jointly. We resolve this chicken-and-egg problem with an EM-type training strategy, where a face autoen- coder is trained jointly with an outlier segmentation net- work. This leads to a synergistic effect, in which the seg- mentation network prevents the face encoder from fitting to the outliers, enhancing the reconstruction quality. The im- proved 3D face reconstruction, in turn, enables the segmen- tation network to better predict the outliers. To resolve the ambiguity between outliers and regions that are difficult to fit, such as eyebrows, we build a statistical prior from syn- thetic data that measures the systematic bias in model fit- ting. Experiments on the NoW testset demonstrate that FO- CUS achieves SOTA 3D face reconstruction performance among all baselines trained without 3D annotation. More- over, our results on CelebA-HQ and AR database show that the segmentation network can localize occluders accurately ∗Denotes same contribution. Codes available at: github.com/unibas-gravis/Occlusion-Robust-MoFA C.Li is funded by the China Scholarship Council (CSC) from the Min- istry of Education of P.R. China. B.Egger was supported by a Post- Doc Mobility Grant, Swiss National Science Foundation P400P2 191110. A.Kortylewski acknowledges support via his Emmy Noether Research Group funded by the German Science Foundation (DFG) under Grant No. 468670075. Sincere gratitude to Tatsuro Koizumi and William A. P. Smith who offered the MoFA re-implementation. Figure 1. FOCUS conducts face reconstruction and outlier seg- mentation jointly under weak supervision. Top to bottom: target images, our reconstruction images, and estimated outlier masks. despite being trained without any segmentation annotation.

1. Introduction Monocular 3D face reconstruction aims at estimating the pose, shape, and albedo of a face, as well as the illumination conditions and camera parameters of the scene. Solving for all these factors from a single image is an ill-posed problem. Model-based face autoencoders [31] overcome this problem through fitting a 3D Morphable Model (3DMM) [1, 9] to a target image. The 3DMM provides prior knowledge about the face albedo and geometry such that 3D face reconstruc- tion from a single image becomes feasible, enabling face autoencoders to set the current state-of-the-art in 3D face reconstruction [5]. The network architectures in the face au- toencoders are devised to enable end-to-end reconstruction and to enhance the speed compared to optimization-based alternatives [19, 41], and sophisticated losses are designed to stabilize the training and to get better performance [5]. A major remaining challenge for face autoencoders is that their performance in in-the-wild environments is still limited by nuisance factors such as model outliers, extreme illumination, and poses. Among those nuisances, model outliers are ubiquitous and inherently difficult to handle be- cause of their wide variety in shape, appearance, and 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. 372 tion. The outliers are a combination of the occlusions that do not belong to the face and the mismatches which are the facial parts but cannot be depicted well by the face model, such as pigmentation and makeup on the texture and wrin- kles on the shape. Fitting to the outliers often distorts the prediction (see Fig. 3) and fitting to the mismatches can- not improve the fitting further due to the limitation of the model. Therefore we propose to only fit the inliers, i.e. the target with the outliers excluded. To prevent distortion caused by model outliers, existing methods often follow a bottom-up approach. For exam- ple, a multi-view shape consistency loss is used as prior to regularize the shape variation of the same face in dif- ferent images [5, 10, 33], or the face symmetry is used to detect occluders [34]. Training the face encoder with dense landmark supervision also imposes strong regulariza- tion [37, 42], while pairs of realistic images and meshes are costly to acquire. Most existing methods apply face [27] or skin [5] segmentation models and subsequently exclude the non-facial regions during reconstruction. These segmenta- tion methods operate in a supervised manner, which suffers from the high cost and efforts for acquiring a great variety of occlusion annotations from in-the-wild images. In this work, we introduce an approach to handle outliers for model-based face reconstruction that is highly robust, without requiring any annotations for skin, occlusions, or dense landmarks. In particular, we propose to train a Face- autOenCoder and oUtlier Segmentation network, abbrevi- ated as FOCUS, in a cooperative manner. To train the seg- mentation network in an unsupervised manner, we exploit the fact that the outliers cannot be well-expressed by the face model to guide the decision-making process of an out- lier segmentation network. Specifically, the discrepancy be- tween the target image and the rendered face image (Fig. 1 1st and 2nd rows) are evaluated by several losses that can serve as a supervision signal by preserving the similarities among the target image, the reconstructed image, and the reconstructed image under the estimated outlier mask. The training process follows the core idea of the Expectation-Maximization (EM) algorithm, by alternating between training the face autoencoder given the currently estimated segmentation mask, and subsequently training the segmentation network based on the current face reconstruc- tion. The EM-like training strategy resolves the chicken- and-egg problem that the outlier segmentation and model fitting are dependent on each other. This leads to a syner- gistic effect, in which the outlier segmentation first guides the face autoencoder to fit image regions that are easy to classify as face regions. The improved face fitting, in turn, enables the segmentation model to refine its prediction. We define in-domain misfits as errors in regions, where a fixed model can explain but constantly not fit well, which are observed in the eyebrows and the lip region. We assumethat such misfits result from the deficiencies of the fitting pipeline. Model-based face autoencoders use image-level losses only, which are highly non-convex and suffer from local optima. Consequently, it is difficult to converge to the globally optimal solution. In this work, we propose to mea- sure and adjust the in-domain misfits with a statistical prior. Our misfit prior learns from synthetic data at which regions these systematic errors occur on average. Subsequently, the learnt prior can be used to counteract these errors for pre- dictions on real data, especially when our FOCUS structure excludes the outliers. Building the prior requires only data generated by a linear face model without any enhancement and no further improvement in landmark detection. We demonstrate the effectiveness of our proposed pipeline by conducting experiments on the NoW testset [29], where we achieve state-of-the-art performance among model-based 3D face methods without 3D supervision. Re- markably, experiments on the CelebA-HQ dataset [20] and the AR database [22] validate that our method is able to predict accurate occlusion masks without requiring any su- pervision during training. In summary, we make the following contributions: 1. We introduce an approach for model-based 3D face reconstruction that is highly robust, without requiring any human skin or occlusion annotation. 2. We propose to compensate for the misfits with an in- domain statistical misfit prior, which is easy to imple- ment and benefits the reconstruction. 3. Our model achieves SOTA performance at self- supervised 3D face reconstruction and provides accu- rate estimates of the facial occlusion masks on in-the- wild images.

Lv_Unbiased_Multiple_Instance_Learning_for_Weakly_Supervised_Video_Anomaly_Detection_CVPR_2023

Abstract Weakly Supervised Video Anomaly Detection (WSVAD) is challenging because the binary anomaly label is only given on the video level, but the output requires snippet- level predictions. So, Multiple Instance Learning (MIL) is prevailing in WSVAD. However, MIL is notoriously known to suffer from many false alarms because the snippet-level detector is easily biased towards the abnormal snippets with simple context, confused by the normality with the same bias, and missing the anomaly with a different pattern. To this end, we propose a new MIL framework: Unbiased MIL (UMIL), to learn unbiased anomaly features that improve WSVAD. At each MIL training iteration, we use the cur- rent detector to divide the samples into two groups with dif- ferent context biases: the most confident abnormal/normal snippets and the rest ambiguous ones. Then, by seeking the invariant features across the two sample groups, we can remove the variant context biases. Extensive exper- iments on benchmarks UCF-Crime and TAD demonstrate the effectiveness of our UMIL. Our code is provided at https://github.com/ktr-hubrt/UMIL.

1. Introduction Video Anomaly Detection (V AD) aims to detect events among video sequences that deviate from expectation, which is widely applied in real-world tasks such as intel- ligent manufacturing [8], TAD surveillance [9,22] and pub- lic security [25, 30]. To learn such a detector, conventional fully-supervised V AD [1] is impractical as the scattered but diverse anomalies require extremely expensive labeling cost. On the other hand, unsupervised V AD [3, 11, 13, 35, 42] by only learning on normal videos to detect open-set anomalies often triggers false alarms, as it is essentially ill- posed to define what is normal and abnormal by giving only *Corresponding author 01 Time01 ScoreExplosion Vandalism Time (a) 01 Time01 ScoreExplosion Vandalism Time (b) Figure 1. Two anomalies of Explosion and Vandalism are illus- trated. Among each video sequence, we use red boxes to highlight the ground-truth anomaly regions as in the first row. The corre- sponding anomaly curves of an MIL-based model are depicted be- low. False alarms and real anomalies are linked to the curves with blue arrows and green arrows respectively. Best viewed in color. normal videos without any prior knowledge. Hence, we are interested in a more practical setting: Weakly Supervised V AD (WSV AD) [12, 43], where only video-level binary la- bels ( i.e., normal vs.abnormal) are available. In WSV AD, each video sequence is partitioned into multiple snippets. Hence, all the snippets are normal in a normal video, and at least one snippet contains the anomaly in an abnormal one. The goal of WSV AD is to train a snippet-level anomaly detector using video-level la- bels. The mainstream method is Multiple Instance Learn- ing (MIL) [22, 30]—multiple instances refer to the snip- pets in each video, and learning is conducted by decreas- ing the predicted anomaly score for each snippet in a nor- mal video, and increasing that only for the snippet with the largest anomaly score in an abnormal video. For example, Figure 1a shows an abnormal video containing an explo- sion scene, and the detector is trained by MIL to increase the anomaly score for the most anomalous explosion snip- pet (green link). However, MIL is easily biased towards the simplest con- text shortcut in a video. We observe in Figure 1a that the 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. 8022 Invariance FeatureContext Feature (a) (b) (c) Failure Cases(a) (c)(b) B BFigure 2. Red: Confident Set, Blue: Ambiguous Set. : Nor- mal sample, ▲: Abnormal sample, Gray instances: Failure cases. The red line denotes the classifier trained under MIL. The invari- ant classifier (black line) can be learned by combining confident snippets learning in MIL (red line) and the ambiguous snippets clustering (blue line). Best viewed in color. tector is biased to smoke, as the pre-explosion snippet with only smoke is also assigned a large anomaly score (blue link). This biased detector can trigger false alarms on smoke snippets without anomaly, e.g., a smoking chimney. More- over, it could also fail in videos with multiple anomalies of different contexts. In Figure 1b, the video records two men vandalizing a car, where only the second one has substantial motions. We notice that the two snippets of them have large differences in the anomaly scores, and only the latter is pre- dicted as an anomaly. This shows that the detector is biased to the drastic motion context while being less sensitive to the subtle vandalism behavior, which is the true anomaly. The root of MIL’s biased predictions lies in its training scheme with biased sample selection. As shown in Fig- ure 2, the bottom-left cluster (denoted as the red ellipse) corresponds to the confident normal snippets, e.g., an empty crossroad or an old man standing in a room, which are either from normal videos as the ground truth or from abnormal videos but visually similar to the ground-truth ones. On the contrary, the top-right cluster denotes the confident abnor- mal ones, which not only contain the true anomaly features (e.g., explosion and vandalism) but also include the context features commonly appearing with anomaly under a context bias ( e.g., smoke and motions). In MIL, the trained detector is dominated by the confident samples, corresponding to the top-right cluster with the abnormal representation and the bottom-left cluster with the normal representation. Hence the learned detector (red line) inevitably captures the con- text bias in the confident samples. Consequently, the biased detector generates ambiguous predictions on snippets with a different context bias (the red line mistakenly crossing the blue points), e.g., smoke but normal (industrial exhaust inFigure 2a), substantial motion but normal (equipment main- tenance in Figure 2b), or subtle motion but abnormal (van- dalizing the rear-view mirror in Figure 2c), leading to the aforementioned failure cases. To this end, we aim to build an unbiased MIL detector by training with both the confident abnormal/normal and the ambiguous ones. Specifically, at each UMIL training iteration, we divide the snippets into two sets using the cur- rent detector: 1) the confident set with abnormal and nor- mal snippets and 2) the ambiguous set with the rest snip- pets, e.g., the two sets are enclosed with red circles and blue circles in Figure 2, respectively. The ambiguous set is grouped into two unsupervised clusters ( e.g., the two blue circles separated by the blue line) to discover the intrinsic difference between normal and abnormal snippets. Then, we seek an invariant binary classifier between the two sets that separate the abnormal/normal in the confident set and the two clusters in the ambiguous one. The rationale of the proposed invariance pursuit is that the snippets in the ambiguous set must have a different context bias from the confident set, otherwise, they will be selected into the same set. Therefore, given a different context but the same true anomaly, the invariant pursuit will turn to the true anomaly (e.g., the black line). Overall, we term our approach as Unbiased MIL (UMIL) . Our contributions are summarized below: • UMIL is a novel WSV AD method that learns an unbiased anomaly detector by pursuing the invariance across the confident and ambiguous snippets with different context biases. • Thanks to the unbiased objective, UMIL is the first WS- V AD method that combines feature fine-tuning and de- tector learning into an end-to-end training scheme. This leads to a more tailored feature representation for V AD. • UMIL is equipped with a fine-grained video partitioning strategy for preserving the subtle anomaly information in video snippets. • These contribute to the improved performance over the current state-of-the-art methods on UCF-Crime [30] ( 1.4%AUC) and TAD [22] ( 3.3%AUC) benchmarks. Note that UMIL brings more than 2% AUC gain com- pared with the MIL baseline on both datasets, which jus- tifies the effectiveness of UMIL.

Li_SCConv_Spatial_and_Channel_Reconstruction_Convolution_for_Feature_Redundancy_CVPR_2023

Abstract Convolutional Neural Networks (CNNs) have achieved remarkable performance in various computer vision tasks but this comes at the cost of tremendous computational resources, partly due to convolutional layers extract- ing redundant features. Recent works either compress well-trained large-scale models or explore well-designed lightweight models. In this paper, we make an attempt to exploit spatial and channel redundancy among features for CNN compression and propose an efficient convolu- tion module, called SCConv (Spatial and Channel recon- struction Convolution), to decrease redundant computing and facilitate representative feature learning. The pro- posed SCConv consists of two units: spatial reconstruction unit (SRU) and channel reconstruction unit (CRU). SRU utilizes a separate-and-reconstruct method to suppress the spatial redundancy while CRU uses a split-transform-and- fuse strategy to diminish the channel redundancy. In addi- tion, SCConv is a plug-and-play architectural unit that can be used to replace standard convolution in various convolu- tional neural networks directly. Experimental results show that SCConv-embedded models are able to achieve better performance by reducing redundant features with signifi- cantly lower complexity and computational costs.

1. Introduction In recent years, convolutional neural networks (CNNs) have obtained widespread applications in computer vision tasks [24] due to its ability in obtaining representative fea- tures. However, such success relies heavily on intensive resources of computation and storage, which poses se- vere challenges to their efficient deployment on resource- constrained environments. Therefore, to address these chal- lenges, various types of model compression strategies and network designs have been explored to improve network ef- *Corresponding author (e-mail: ywen@cs.ecnu.edu.cn)ficiency [1, 2, 26]. The former includes network pruning, weight quantization, low-rank factorization, and knowledge distillation. To be specific, network pruning [17,22,30] is a straightforward way to prune the uncritical neuron connec- tions from an existing learned big model to make it thinner. Weight quantization [9] mainly focuses on converting net- work weights from floating-point types to integer ones to save computation resources. Low-rank factorization [5] ap- plies the matrix decomposition techniques to estimate the informative parameters. Knowledge distillation [11, 34] generates small student networks with the guidance of a well-trained big teacher network. The common part of these compression techniques is that they have been regarded as post-processing steps, thus their performance is usually up- per bounded by the given initial model. Meanwhile, the ac- curacy of these methods drastically drops while achieving a high compression rate. Network design is another alternative way, which aims at reducing the inherent redundancy in dense model param- eters and further developing a lightweight network model. For example, ResNet [10] and DenseNet [14] utilize an effi- cient shortcut connection to improve the network topology, which connects all preceding feature maps to diminish the redundant parameters. ResNeXt [31] replaces traditional convolutions with sparsely connected group convolutions to reduce inter-channel connectivity. Networks like Xcep- tion [4], MobileNet [12] and MobileNeXt [35] disentan- gle standard convolution into depth-wise convolution and point-wise convolution to further decrease the connection density between channels. MicroNet [19] adopts micro- factorized convolution to handle extremely low FLOPs by integrating sparse connectivity into convolution. In addi- tion, EfficientNet [27] learns to automatically search opti- mal network architectures to lower the redundancy in dense model parameters. Moreover, in CNN architecture design, bottleneck struc- ture has been well adopted, in which 3×3convolutional layers account for a majority of the model parameters and FLOPs. Therefore various efficient convolutional opera- 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.

Pietrantoni_SegLoc_Learning_Segmentation-Based_Representations_for_Privacy-Preserving_Visual_Localization_CVPR_2023

Abstract Inspired by properties of semantic segmentation, in this paper we investigate how to leverage robust image segmen- tation in the context of privacy-preserving visual localiza- tion. We propose a new localization framework, SegLoc, that leverages image segmentation to create robust, com- pact, and privacy-preserving scene representations, i.e., 3D maps. We build upon the correspondence-supervised, fine- grained segmentation approach from [42], making it more robust by learning a set of cluster labels with discriminative clustering, additional consistency regularization terms and we jointly learn a global image representation along with a dense local representation. In our localization pipeline, the former will be used for retrieving the most similar im- ages, the latter to refine the retrieved poses by minimizing the label inconsistency between the 3D points of the map and their projection onto the query image. In various ex- periments, we show that our proposed representation al- lows to achieve (close-to) state-of-the-art pose estimation results while only using a compact 3D map that does not contain enough information about the original images for an attacker to reconstruct personal information.

1. Introduction Visual localization is the problem of estimating the pre- cise camera pose – position and orientation – from which the image was taken in a known scene. It is a core compo- nent of systems such as self-driving cars [31], autonomous robots [49], and mixed-reality applications [4, 53]. Traditionally, visual localization algorithms rely on a 3D scene representation of the target area, which can be a 3D point cloud map [29, 34, 35, 45, 46, 66, 68, 69, 73, 79], e.g., from Structure-from-Motion (SfM), or a learned 3D repre- sentation [9,10,14,37,38,71,76]. The representation is typ- ically derived from reference images with known camera poses. Depending on the application scenario, these maps (𝑅,𝑇)Pose refinement through label alignment(𝑅!,𝑇!) Queryimage SegLoc Image retrievalGlobal descriptorSegmentation+Labelled 3D mapFigure 1. The SegLoc localization pipeline: Our model jointly creates a robust global descriptor used to retrieve an initial pose (R0,T0)and dense local representations used to obtain the re- fined pose (R,T)by maximizing the label consistency between the reprojected 3D points and the query image. need to be stored in the cloud, which raises important ques- tions about memory consumption andprivacy preserva- tion. It is possible to reconstruct images from maps that contain local image features [62], amongst the most widely used for scene representation. To tackle the above challenges that feature-based ap- proaches may face, inspired by semantic-based [48,82] and segmentation-based [42] approaches, we propose a visual localization pipeline where robust segmentations are used as the sole cue for localization, yielding reduced storage requirements (compared to using local features) while in- creasing privacy-preservation. Our proposed localization pipeline, called SegLoc, follows standard structure based- localization pipelines [34, 66] that represent the scene via a 3D model: first, image retrieval based on a compact im- age representation is used to coarsely localize a query im- age. Given such an initial pose estimate, the camera pose is refined by aligning the query image to the 3D map. Con- trary to prior work that is based on extracting features di- rectly from images, we derive a more abstract representa- tion in the form of a robust dense segmentation based on 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. 15380 set of clusters learned in a self-supervised manner. As illus- trated in Figure 1, we use this segmentation to both extract a global descriptor for image retrieval and for pose refine- ment. The pose is refined by maximizing the label consis- tency between the predictions in the query image and a set of labeled 3D points in the scene. Such an approach has multiple advantages. First, our model is able to learn representations which are robust to seasonal or appearance changes . Similar to semantic seg- mentations, which are invariant to viewing conditions as the semantic meaning of regions do not change, our represen- tation is trained such that the same 3D point is mapped to the same label regardless of viewing conditions. Second, it results in low storage requirements , as instead of storing high-dimensional feature descriptors, for each 3D point we only keep its label. Finally, it allows privacy-preserving vi- sual localization [15,22,28,78], as it creates a non-injective mapping from multiple images showing similar objects with different appearances to similar labels. While, ensuring user privacy comes at the cost of reduced pose accuracy [19,98], our method comes close to state-of-the-art results with a better accuracy vs. memory vs. privacy trade-off. To summarize, our first contribution is a new localiza- tion framework, called SegLoc , that extends the idea [41, 42] of learning robust fine-grained image segmentations in a self-supervised manner. To that end, we leverage dis- criminative clustering while putting more emphasis on rep- resentation learning. Furthermore, we derive a full local- ization pipeline, where our model jointly learns global im- age representation to retrieve images for pose initialization, and dense local representations for building a compact 3D map – an order of magnitude smaller compared to feature- based approaches – and to perform privacy-preserving pose refinement. As a second contribution , we draw a con- nection between segmentation-based representations and privacy-preserving localization, opening up viable alter- natives to keypoint-based methods within the accuracy- privacy-memory trade-off. We evaluate our approach in multiple indoor and outdoor environments while quantita- tively measuring privacy through detailed experiments.

Mou_Large-Capacity_and_Flexible_Video_Steganography_via_Invertible_Neural_Network_CVPR_2023

Abstract Video steganography is the art of unobtrusively conceal- ing secret data in a cover video and then recovering the secret data through a decoding protocol at the receiver end. Although several attempts have been made, most of them are limited to low-capacity and fixed steganography. To rectify these weaknesses, we propose a Large-capacity and Flexible Video Steganography Network (LF-VSN) in this paper. For large-capacity, we present a reversible pipeline to perform multiple videos hiding and recovering through a single invertible neural network (INN). Our method can hide/recover 7 secret videos in/from 1 cover video with promising performance. For flexibility, we propose a key- controllable scheme, enabling different receivers to recover particular secret videos from the same cover video through specific keys. Moreover, we further improve the flexibility by proposing a scalable strategy in multiple videos hid- ing, which can hide variable numbers of secret videos in a cover video with a single model and a single training session. Extensive experiments demonstrate that with the significant improvement of the video steganography perfor- mance, our proposed LF-VSN has high security, large hid- ing capacity, and flexibility. The source code is available at https://github.com/MC-E/LF-VSN .

1. Introduction Steganography [10] is the technology of hiding some se- cret data into an inconspicuous cover medium to generate a stego output, which only allows the authorized receiver to recover the secret information. Unauthorized people can only access the content of the plain cover medium, and hard ∗Corresponding author . This work was supported by the King Abdul- lah University of Science and Technology (KAUST) Office of Sponsored Research through the Visual Computing Center (VCC) funding, SDAIA- KAUST Center of Excellence in Data Science and Artificial Intelligence, and Shenzhen Research Project JCYJ20220531093215035.to detect the existence of secret data. In the current digital world, image and video are commonly used covers, widely applied in digital communication [27], copyright protec- tion [36], information certification [31], e-commerce [26], and many other practical fields [10, 12]. Traditional video steganography methods usually hide messages in the spatial domain or transform domain by manual design. Video steganography in the spatial domain means embedding is done directly to the pixel values of video frames. Least significant bits (LSB) [8,45] is the most well-known spatial-domain method, replacing the nleast significant bits of the cover image with the most significant nbits of the secret data. Many researchers have used LSB replacement [6] and LSB matching [34] for video steganog- raphy. The transform-domain hiding [5, 17, 39] is done by modifying certain frequency coefficients of the transformed frames. For instance, [44] proposed a video steganogra- phy technique by manipulating the quantized coefficients of DCT (Discrete Cosine Transformation). [9] proposed to compare the DWT (Discrete Wavelet Transformation) co- efficients of the secret image and the cover video for hid- ing. However, these traditional methods have low hiding capacity and invisibility, easily being cracked by steganaly- sis methods [15, 28, 33]. Recently, some deep-learning methods were proposed to improve the hiding capacity and performance. Early works are presented in image steganography. Baluja [3, 4] pro- posed the first deep-learning method to hide a full-size im- age into another image. Recently, [21,32] proposed design- ing the steganography model as an invertible neural network (INN) [13,14] to perform image hiding and recovering with a single model. For video steganography, Khare et al. [22] first utilized back propagation neural networks to improve the performance of the LSB-based scheme. [43] is the first deep-learning method to hide a video into another video. Unfortunately, it simply aims to hide the residual across ad- jacent frames in a frame-by-frame manner, and it requires several separate steps to complete the video hiding and 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. 22606 Figure 1. Illustration of our large-capacity and flexible video steganography network (LF-VSN). Our LF-VSN reversibly solves multiple videos hiding and recovering with a single model and the same parameters. It has large-capacity, key-controllable and scalable advantages. covering. [35] utilize 3D-CNN to explore the temporal cor- relation in video hiding. However, it utilizes two separated 3D UNet to perform hiding and recovering, and it has high model complexity ( 367.2million parameters). While video steganography has achieved impressive success in terms of hiding capacity to hide a full-size video, the more challeng- ing multiple videos hiding has hardly been studied. Also, the steganography pipeline is rigid. In this paper, we study the large-capacity and flexible video steganography, as shown in Fig. 1. Concretely, we propose a reversible video steganography pipeline, achiev- ing large capacity to hide/recover multiple secret videos in/from a cover video. At the same time, our model complexity is also attractive by combining several weight- sharing designs. The flexibility of our method is twofold. First, we propose a key-controllable scheme, enabling dif- ferent receivers to recover particular secret videos with spe- cific keys. Second, we propose a scalable strategy, which can hide variable numbers of secret videos into a cover video with a single model and a single training session. To summarize, this work has the following contributions: • We propose a large-capacity video steganography method, which can hide/recover multiple ( up to 7 ) se- cret videos in/from a cover video. Our hiding and re- covering are fully reversible via a single INN. • We propose a key-controllable scheme with which dif- ferent receivers can recover particular secret videos from the same cover video via specific keys. • We propose a scalable embedding module, utilizing a single model and a single training session to satisfy different requirements for the number of secret videos hidden in a cover video. • Extensive experiments demonstrate that our proposedmethod achieves state-of-the-art performance with large hiding capacity and flexibility.

Lo_Spatio-Temporal_Pixel-Level_Contrastive_Learning-Based_Source-Free_Domain_Adaptation_for_Video_Semantic_CVPR_2023

Abstract Unsupervised Domain Adaptation (UDA) of semantic segmentation transfers labeled source knowledge to an un- labeled target domain by relying on accessing both the source and target data. However, the access to source data is often restricted or infeasible in real-world scenar- ios. Under the source data restrictive circumstances, UDA is less practical. To address this, recent works have ex- plored solutions under the Source-Free Domain Adaptation (SFDA) setup, which aims to adapt a source-trained model to the target domain without accessing source data. Still, existing SFDA approaches use only image-level informa- tion for adaptation, making them sub-optimal in video ap- plications. This paper studies SFDA for Video Semantic Segmentation (VSS), where temporal information is lever- aged to address video adaptation. Specifically, we pro- pose Spatio-Temporal Pixel-Level (STPL) contrastive learn- ing, a novel method that takes full advantage of spatio- temporal information to tackle the absence of source data better. STPL explicitly learns semantic correlations among pixels in the spatio-temporal space, providing strong self- supervision for adaptation to the unlabeled target domain. Extensive experiments show that STPL achieves state-of- the-art performance on VSS benchmarks compared to cur- rent UDA and SFDA approaches. Code is available at: https://github.com/shaoyuanlo/STPL

1. Introduction The availability of large amounts of labeled data has made it possible for various deep networks to achieve re- markable performance on Image Semantic Segmentation (ISS) [2, 4, 30]. However, these deep networks often general- ize poorly on target data from a new unlabeled domain that is visually distinct from the source training data. Unsupervised Domain Adaptation (UDA) attempts to mitigate this domain shift problem by using both the labeled source data and un- *This work was mostly done when S.-Y . Lo was an intern at Amazon. Figure 1. Comparison of VSS accuracy. Video-based UDA meth- ods [12, 38, 49] outperform image-based UDA methods [33, 51], showing the importance of video-based strategies for the VSS task. Image-based SFDA methods [16, 39] perform lower than the UDA methods, which shows the difficulty of the more restricted SFDA setting. The proposed STPL, even with SFDA, achieves the best accuracy and locates at the top-right corner of the chart (i.e., more restriction, but higher accuracy). labeled target data to train a model transferring the source knowledge to the target domain [11, 12, 31, 32, 38, 41]. UDA is effective but relies on the assumption that both source and target data are available during adaptation. In real-world scenarios, the access to source data is often restricted (e.g., data privacy, commercial proprietary) or infeasible (e.g., data transmission efficiency, portability). Hence, under these source data restrictive circumstances, UDA approaches are less practical. To deal with these issues, the Source-Free Domain Adap- tation (SFDA) setup, also referred to as Unsupervised Model Adaptation (UMA), has been recently introduced in the lit- erature [6, 26, 27, 52]. SFDA aims to use a source-trained 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. 10534 model (i.e., a model trained on labeled source data) and adapt it to an unlabeled target domain without requiring access to the source data. More precisely, under the SFDA formula- tion, given a source-trained model and an unlabeled target dataset, the goal is to transfer the learned source knowledge to the target domain. In addition to alleviating data privacy or proprietary concerns, SFDA makes data transmission much more efficient. For example, a source-trained model ( 0.1 - 1.0 GB) is usually much smaller than a source dataset (  10 - 100 GB). If one is adapting a model from a large-scale cloud center to a new edge device that has data with different domains, the source-trained model is far more portable and transmission-efficient than the source dataset. Under SFDA, label supervision is not available. Most SFDA studies adopt pseudo-supervision or self-supervision techniques to adapt the source-trained model to the target domain [16, 39]. However, they consider only image-level information for model adaptation. In many real-world seman- tic segmentation applications (autonomous driving, safety surveillance, etc.), we have to deal with temporal data such as streams of images or videos. Supervised approaches that use temporal information have been successful for Video Seman- tic Segmentation (VSS), which predicts pixel-level semantics for each video frame [19, 22, 28, 46]. Recently, video-based UDA strategies have also been developed and yielded better performance than image-based UDA on VSS [12, 38, 49]. This motivates us to propose a novel SFDA method for VSS, leveraging temporal information to tackle the absence of source data better. In particular, we find that current image-based SFDA approaches suffer from sub-optimal per- formance when applied to VSS (see Figure 1). To the best of our knowledge, this is the first work to explore video-based SFDA solutions. In this paper, we propose a novel spatio-temporal SFDA method namely Spatio-Temporal Pixel-Level (STPL) Con- trastive Learning (CL), which takes full advantage of both spatial and temporal information for adapting VSS mod- els. STPL consists of two main stages. (1) Spatio-temporal feature extraction: First, given a target video sequence in- put, STPL fuses the RGB and optical flow modalities to extract spatio-temporal features from the video. Meanwhile, it performs cross-frame augmentation via randomized spatial transformations to generate an augmented video sequence, then extracts augmented spatio-temporal features. (2) Pixel- level contrastive learning: Next, STPL optimizes a pixel- level contrastive loss between the original and augmented spatio-temporal feature representations. This objective en- forces representations to be compact for same-class pixels across both the spatial and temporal dimensions. With these designs, STPL explicitly learns semantic corre- lations among pixels in the spatio-temporal space, providing strong self-supervision for adaptation to an unlabeled tar- get domain. Furthermore, we demonstrate that STPL is anon-trivial unified spatio-temporal framework. Specifically, Spatial-only CL andTemporal-only CL are special cases of STPL, and STPL is better than a na ¨ıve combination of them. Extensive experiments demonstrate the superiority of STPL over various baselines, including the image-based SFDA as well as image- and video-based UDA approaches that rely on source data (see Figure 1). The key contributions of this work are summarized as follows: •We propose a novel SFDA method for VSS. To the best of our knowledge, this is the first work to explore video-based SFDA solutions. •We propose a novel CL method, namely STPL, which explicitly learns semantic correlations among pixels in the spatio-temporal space, providing strong self-supervision for adaptation to an unlabeled target domain. •We conduct extensive experiments and show that STPL provides a better solution compared to the existing image- based SFDA methods as well as image- and video-based UDA methods for the given problem formulation.

Mei_Deep_Polarization_Reconstruction_With_PDAVIS_Events_CVPR_2023

Abstract The polarization event camera PDAVIS is a novel bio- inspired neuromorphic vision sensor that reports both con- ventional polarization frames and asynchronous, continu- ously per-pixel polarization brightness changes (polariza- tion events) with fast temporal resolution andlarge dy- namic range . A deep neural network method (Polariza- tion FireNet) was previously developed to reconstruct the polarization angle and degree from polarization events for bridging the gap between the polarization event camera and mainstream computer vision. However, Polarization FireNet applies a network pre-trained for normal event- based frame reconstruction independently on each of four channels of polarization events from four linear polariza- tion angles, which ignores the correlations between chan- nels and inevitably introduces content inconsistency be- tween the four reconstructed frames, resulting in unsatisfac- tory polarization reconstruction performance. In this work, we strive to train an effective, yet efficient, DNN model thatdirectly outputs polarization from the input raw polariza- tion events. To this end, we constructed the first large- scale event-to-polarization dataset, which we subsequently employed to train our events-to-polarization network E2P . E2P extracts rich polarization patterns from input polariza- tion events and enhances features through cross-modality context integration. We demonstrate that E2P outperforms Polarization FireNet by a significant margin with no addi- tional computing cost. Experimental results also show that E2P produces more accurate measurement of polarization than the PDAVIS frames in challenging fast and high dy- namic range scenes. Code and data are publicly available at:https://github.com/SensorsINI/e2p .

1. Introduction Visual information is encoded in light by intensity, color, and polarization [12]. Polarization is a property of trans- verse light waves that specifies the geometric orientation 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. 22149 the oscillations (which can be described by the Angle of Linear Polarization ( AoLP ) and the Degree of Linear Polar- ization ( DoLP )), providing strong vision cues and enabling solutions to challenging problems in medical [27], under- water [34], and remote sensing [53] applications. Existing polarization digital cameras capture synchronous polariza- tion frames with a linear photo response [14], while biologi- cal eyes tend to perceive asynchronous and sparse data with a compressed non-linear response [12]. Inspired by the mantis shrimp visual system [29], the novel neuromorphic vision sensor called Polarization Dy- namic and Active pixel VIsion Sensor ( PDA VIS ) illustrated in Figure 1 was developed to concurrently record a high- frequency stream of asynchronous polarization brightness change events under four polarization angles ( i.e.,0◦,45◦, 90◦, and 135◦) over a wide range of illumination. PDA VIS also outputs low-frequency synchronous frames like con- ventional polarization cameras [15]. Even though the stream of polarization events has ad- vantages of low latency and HDR, it is not friendly to hu- man observation and traditional computer vision due to the sparse, irregular, and unstructured properties. To better ex- ploit the advantages of PDA VIS, an intuitive solution is to reconstruct polarization from polarization events, which can bridge off-the-shelf frame-based algorithms and PDA VIS. Gruev et al. [15] proposed the Polarization FireNet, which first runs the FireNet [41] pre-trained for normal event- based intensity frame reconstruction on each of four types of polarization events under four different polarization an- gles, and then computes the polarization from four recon- structed intensity frames via mathematical formulas. Since this method treats four polarization angle channels indepen- dently, the correlation between channels is ignored and in- consistency between the four reconstructed frames hinders accurate measurement of polarization. In this work, we make the first attempt to train an accurate yet efficient DNN model tailored for event-to- polarization reconstruction. We approach this twofold. First, we construct the first large-scale event-to-polarization synthetic-real mixed dataset, dubbed Events to Polariza- tion Dataset ( E2PD ), which contains 5 billion polarization events and corresponding 133 thousand polarization video frames. The diversity and practicality of E2PD are ensured by including diverse real-world road scenes under differ- ent weather conditions (rainy and sunny) in different cities. Second, we design an E2P network that consists of three branches to reconstruct intensity, AoLP, and DoLP, respec- tively, from the raw polarization events directly. E2P is built on two key modules: (i) a Rich Polarization Pattern Per- ception ( RPPP ) module that effectively harvests features from raw polarization events and (ii) a Cross-Modality At- tention Enhancement ( CMAE ) module that explores cross- modality contextual cues for feature enhancement.We perform extensive validation experiments to demon- strate the efficacy of our method and show that the network trained on our E2PD is more accurate than all previously reported PDA VIS methods, and produces more accurate po- larization compared with polarization computed from the PDA VIS frames in challenging scenes ( e.g., Figure 1). In summary, our contributions are: 1. the first attempt to solve the event-to-polarization problem using an end-to-end trained deep neural net- work with polarization events as input, intensity, AoLP and DoLP as outputs; 2. a new and unique large-scale event-to-polarization dataset containing both synthetic and real data; and 3. a novel network that perceives rich polarization pat- terns from raw polarization events and enhances fea- tures via a cross-modality attention mechanism.

Ma_CAT_LoCalization_and_IdentificAtion_Cascade_Detection_Transformer_for_Open-World_Object_CVPR_2023

Abstract Open-world object detection (OWOD), as a more gen- eral and challenging goal, requires the model trained from data on known objects to detect both known and unknown objects and incrementally learn to identify these unknown objects. The existing works which employ standard de- tection framework and fixed pseudo-labelling mechanism (PLM) have the following problems: (𝑖)The inclusion of de- tecting unknown objects substantially reduces the model’s ability to detect known ones. (𝑖𝑖)The PLM does not ad- equately utilize the priori knowledge of inputs. (𝑖𝑖𝑖)The fixed selection manner of PLM cannot guarantee that the model is trained in the right direction. We observe that hu- mans subconsciously prefer to focus on all foreground ob- jects and then identify each one in detail, rather than lo- calize and identify a single object simultaneously, for al- leviating the confusion. This motivates us to propose a novel solution called CAT: Lo Calization and Identific Ation Cascade Detection Transformer which decouples the detec- tion process via the shared decoder inthe cascade decod- ing way . In the meanwhile, we propose the self-adaptive pseudo-labelling mechanism which combines the model- driven with input-driven PLM and self-adaptively generates robust pseudo-labels for unknown objects, significantly im- proving the ability of CAT to retrieve unknown objects. Ex- periments on two benchmarks, 𝑖.𝑒., MS-COCO and PAS- CAL VOC, show that our model outperforms the state-of- the-art methods. The code is publicly available at https: //github.com/xiaomabufei/CAT .

1. Introduction Open-world object detection (OWOD) is a more prac- tical detection problem in computer vision, making artifi- *Equal contribution. †Corresponding author. BearFrogFlower SquirrelUnknown CatBeeUnknown UnknownUnknown UnknownFigure 1. When faced with new scenes in open world, humans sub- consciously focus on all foreground objects and then identify them in detail in order to alleviate the confusion between the known and unknown objects and get a clear view. Motivated by this, our CAT utilizes the shared decoder to decouple the localization and iden- tification process in the cascade decoding way, where the former decoding process is used for localization and the latter for identi- fication. cial intelligence (AI) smarter to face more difficulties in real scenes. Within the OWOD paradigm, the model’s life-span is pushed by iterative learning process. At each episode, the model trained only by known objects needs to detect known objects while simultaneously localizing unknown objects and identifying them into the unknown class. Human an- notators then label a few of these tagged unknown classes of interest gradually. The model given these newly-added annotations will continue to incrementally update its knowl- edge without retraining from scratch. Recently, Joseph et al. [21] proposed an open-world ob- ject detector, ORE, based on the two-stage Faster R-CNN [38] pipeline. ORE utilized an auto-labelling step to obtain pseudo-unknowns for training model to detect unknown ob- jects and learned an energy-based binary classifier to dis- tinguish the unknown class from known classes. However, its success largely relied on a held-out validation set which 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. 19681 was leveraged to estimate the distribution of unknown ob- jects in the energy-based classifier. Then, several methods [29, 43–45] attempted to extend ORE and achieved some success. To alleviate the problems in ORE, Gupta et al. [17] proposed to use the detection transformer [4,46] for OWOD in a justifiable way and directly leveraged the framework of DDETR [46]. In addition, they proposed an attention- driven PLM which selected pseudo labels for unknown ob- jects according to the attention scores. For the existing works, we find the following hindering problems.(𝑖)Owing to the inclusion of detecting unknown objects, the model’s ability to detect known objects substan- tially drops. To alleviate the confusion between known and unknown objects, humans prefer to dismantle the process of open-world object detection rather than parallelly localize and identify open-world objects like most standard detec- tion models.(𝑖𝑖)To the best of our knowledge, in the exist- ing OWOD PLM, models leverage the learning process for known objects to guide the generation of pseudo l

Lu_Neuron_Structure_Modeling_for_Generalizable_Remote_Physiological_Measurement_CVPR_2023

Abstract Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily af- fected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we sys- tematically address the domain shift problem in the rPPG measurement task. We show that most domain generaliza- tion methods do not work well in this problem, as domain la- bels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron STructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Be- sides, NEST can also enrich and enhance domain invari- ant features across multi-domain. We create and bench- mark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings. The codes are available at https://github.com/LuPaoPao/NEST.

1. Introduction Physiological signals such as heart rate (HR), and heart rate variability (HRV), respiration frequency (RF) are im- portant body indicators that serve not only as vital signs but also track the level of sympathetic activation [17, 33, 54]. Traditional physiological measurements, such as electrocar- diograms, heart rate bands, and finger clip devices, have high accuracy. However, they are costly, intrusive, and un- comfortable to wear for a long time. Remote photoplethys- mography (rPPG) can extract blood volume pulse (BVP) *Corresponding author. VIPL-HR UBFC-rPPG V4V BUAA PURE (a) Typical samples from different datasets GREEN[58] CHROM[8] POS[62] DeepPhys[6] TS-CAN[24] 5 6 7 8 9 109.18 8.92 8.04 8.42 8.25 7.91 Rhythmnet[35] 8.08 7.97AD[11] GroupDRO[37]Traditional DL-based DG-based RMSE↓ (bpm)Ours 6.79Ours (b) The performance of different methods on DG protocolFigure 1. (a) Typical samples from different publicly rPPG datasets: VIPL-HR [36], V4V [47], UBFC-rPPG [1], BUAA [68], PURE [52]. (b) The performance of different methods on DG pro- tocol (test on the UBFC-rPPG dataset with training on the VIPL, V4V , PURE, and BUAA). from face video, which analyzes the periodic changes of the light absorption of the skin caused by heartbeats. Then var- ious physiological indicators (such as HR and HRV) can be calculated based on BVP signals [37, 69, 70]. With non- intrusion and convenience, the rPPG-based physiological measurement method can only use an ordinary camera to monitor physiological indicators and gradually become a re- search hotspot in the computer vision field [7,28,32,36,44, 49]. Traditional rPPG measurement methods include signal blind decomposition [21,34,42] and color space transforma- tion [9, 59, 63]. These approaches rely on heartbeat-related 1 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. 18589 statistical information, only applicable in constrained envi- ronments. In recent years, deep learning (DL) based ap- proaches [6, 19, 25, 28, 36, 44, 50, 69, 70] have shown their great potentials in rPPG measurement. By learning dedi- cated rPPG feature representation, these methods achieve promising performance in much more complicated environ- ments [1, 36, 47, 52, 68]. However, deep learning methods suffer from significant performance degradation when applied in real-world sce- narios. This is because most training data are captured in lab environments with limited environmental variations. With domain shifts (e.g., different illumination, camera param- eters, motions, etc.), these models may struggle to gener- alize for the unseen testing domain. To validate this, we conduct a cross-dataset evaluation shown in Fig. 1(b). As shown, all DL-based methods do not work well in this eval- uation. Furthermore, it is worth noting that DeepPhys [6] and TS-CAN [25] even perform inferior to traditional ap- proach POS [63]. To improve the performance on the unseen domains, one common practice is to incorporate Domain Generalization (DG) approaches, e.g., encouraging intermediate features to be domain-invariant [12, 18, 30, 38, 57, 62]. However, as shown in Fig. 1 (b) (AD [12] and GroupDRO [38]), the improvements are still quite limited. One reason is that existing DG methods assume that domain labels can be clearly defined (e.g., by data source). Unfortunately, differ- ent background scenes, acquisition devices, or even iden- tities could cause very different data distributions in rPPG measurement tasks, and such distribution discrepancy may exist within or cross datasets. In this case, explicitly defin- ing domains is very difficult, and simply treating one dataset as one domain may lead to inferior performance. This is termed as agnostic domain generalization . Besides this, as physiological cues are usually much more subtle than the various noise, the model may overfit in the source domain with the limited training data. In this paper, we propose the NEural STructure model- ing (NEST), a principled solution to the abovementioned problems. The main idea of NEST is to narrow the under- optimized and redundant feature space, align domain invari- ant features, and enrich discriminative features. Our intu- ition is as follows. Neural structure refers to the channel ac- tivation degree in each convolution layer, which reveals the discriminative feature combination for the specific sample. As the limited variation in a certain restricted domain, there are some spaces that the model seldomly is optimized. Out- of-distribution (OOD) samples may cause abnormal activa- tion in these spaces, which may lead to performance de- generation. Therefore, we regularize the neural structure to encourage the model to be more well-conditioned to avoid abnormal activation caused by OOD samples. Specif- ically, we propose the NEural STructure Coverage Maxi-mization (NEST-CM) that encourages all neural spaces to be optimized during training, reducing the chance of ab- normal activation during testing. Secondly, we propose the NEural STructure Targeted Alignment (NEST-TA) that en- courage network suppresses domain variant feature by com- paring the samples with the similar physiological informa- tion. Thirdly, we propose the NEural STructure Diversity Maximization (NEST-DM) to enrich discriminative features against unseen noise. It should be noted that our approach does not rely on domain labels, which is more applicable in the rPPG measurement task. To summarize, our contribu- tions are listed as follows: 1. We are the first to study the domain shift problem in rPPG measurement, which introduces a new challenge, agnostic domain generalization. 2. We propose the NEural STructure modeling to alle- viate domain shift, which is performed by narrowing the under-optimized feature space, and enhancing and enrich- ing domain invariant features. 3. We establish a large-scale domain generalization (DG) benchmark for rPPG measurement, which is the first DG protocol in this task. Extensive experiments in this dataset show the superiority of our approach.

Mai_DualRel_Semi-Supervised_Mitochondria_Segmentation_From_a_Prototype_Perspective_CVPR_2023

Abstract Automatic mitochondria segmentation enjoys great pop- ularity with the development of deep learning. However, existing methods rely heavily on the labor-intensive manual gathering by experienced domain experts. And naively ap- plying semi-supervised segmentation methods in the natural image field to mitigate the labeling cost is undesirable. In this work, we analyze the gap between mitochondrial im- ages and natural images and rethink how to achieve effec- tive semi-supervised mitochondria segmentation, from the perspective of reliable prototype-level supervision. We pro- pose a novel end-to-end dual-reliable (DualRel) network, including a reliable pixel aggregation module and a reliable prototype selection module. The proposed DualRel enjoys several merits. First, to learn the prototypes well without any explicit supervision, we carefully design the referential correlation to rectify the direct pairwise correlation. Sec- ond, the reliable prototype selection module is responsible for further evaluating the reliability of prototypes in con- structing prototype-level consistency regularization. Exten- sive experimental results on three challenging benchmarks demonstrate that our method performs favorably against state-of-the-art semi-supervised segmentation methods. Im- portantly, with extremely few samples used for training, Du- alRel is also on par with current state-of-the-art fully super- vised methods.

1. Introduction Mitochondria, as one of the crucial organelles, are the primary energy providers for cell activities and are essen- tial for metabolism. Quantification of mitochondrial mor- phology can not only promote basic scientific research ( e.g., cellular physiology [1, 5]), but also provide new insight for clinical diagnosis ( e.g., neurodegenerative diseases [20] and diabetes [24]). Recently, with the development of deep learning, semantic segmentation [2, 14, 18, 27, 30, 33] en- ables in-depth exploration of mitochondrial morphology *Equal contribution †Corresponding author Correlation of and : 0.8 Referential Correlations Similarity : 0.8 Foreground prototype Background pixelReferential Correlations Similarity : 0.2 Foreground pixel Pixel -Reference CorrelationPixel -Prototype CorrelationCorrelation of and : 0.9 Reference Points Prototype -Reference Correlation (b) (c)(a)Confusion Density: 0.06 Confusion Density: 0.21Figure 1. Illustration of our motivation. (a) shows the confusion map and density ( i.e., the expected inverse confidence per pixel) of mitochondrial and natural images. (b) shows the unreliability caused by direct pairwise prototype-pixel correlation that is condi- tioned only on visual similarity. (c) shows how to construct pixel- reference correlation to rectify the direct pairwise correlation in a referential correlation manner. from high-resolution electron microscopy (EM) images and make conspicuous achievements. However, their flexibil- ity and scalability are limited in the actual deployment be- cause the numerous cluttered irrelevant organelles that re- quire labor-intensive manual discrimination and gathering by experienced domain experts [10, 21]. Therefore, we be- gin to turn attention to semi-supervised segmentation with the assumption that enormous unlabeled data is accessible, aiming to alleviate the data-hungry issue. Semi-supervised segmentation enjoys great popularity in the field of natural images, and representative works such as CPS [3], which imposes pixel-level consistency regulariza- tion and establishes state-of-the-art performance. It natu- rally comes into mind to directly apply a CPS-like method to semi-supervised mitochondria segmentation. However, there exist a large gap between mitochondrial and natu- ral images. As shown in Fig. 1 (a), we observe that the confusion density ( i.e., the expected inverse confidence per pixel) in mitochondrial images significantly surpasses coun- terpart in natural images, implying that directly employing 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. 19617 pixel-level consistency regularization as supervision signals on mitochondrial images will inevitably increase the risk of unreliability. The most intuitive example is that there exist considerable boundary regions in mitochondrial im- ages, and the segmentation network is naturally equivocal for these regions, as proven in [15]. In this case, some rel- atively small mitochondria are easily overwhelmed by this ambiguity, leading to sub-optimal results. In order to seek more reliable supervision signals to al- leviate the undependable problem caused by pixel-level su- pervision, we draw inspiration from the inbuilt resistance to noisy pixels of prototypes and construct more robust and reliable prototype-level supervision . To achieve this goal, two issues need to be considered. (1) Unreliable pixels. Considering the cluttered background caused by under/overexposure and out-of-focus problems during EM imaging, the prototype inevitably absorbs unreliable pixels (i.e., heterogeneous semantic clues) during the interaction of corresponding pixels with a suitable pattern. We ar- gue that directly forcing pairwise prototype-pixel correla- tion is primarily at blame. As shown in Fig. 1 (b), due to the foreground-background ambiguity, the foreground pro- totype f1is erroneously closer to p2located in the back- ground than counterpart point p1with similar pattern situ- ated in the foreground . Therefore, it is highly desirable to suppress the unreliable pixels caused by the direct pairwise prototype-pixel correlation that is only conditioned on vi- sual similarity during prototype learning process. (2) Unre- liable prototypes. Intuitively, not all prototypes are equiva- lent for building prototype-level consistency regularization. For example, for a prototype that focuses on mitochondrial boundary patterns, the inherent unreliability of the pixels belonging to these patterns, as discussed above, will also taint the purity of this prototype with equivocality. There- fore, the prototype-level supervision signals should be fur- ther optimized to guarantee that the true reliable prototypes enjoy higher weights. To mitigate the above issues, we rethink how to achieve effective consistency regularization for semi-supervised mi- tochondria segmentation, from the perspective of reliable prototype-level supervision. We propose a Dual-Rel iable (DualRel) network including a reliable pixel aggregation module and a reliable prototype selection module. In the reliable pixel aggregation module (RPiA), to learn the prototypes well without any explicit supervision, we care- fully design the referential correlation to rectify the direct pairwise correlation, enabling the prototype absorb counter- part reliable pixels with the same semantic pattern during the interaction with the pixels. The main idea is, for each pixel/prototype, we can obtain the referential correlation (i.e., a likehood vector) by comparing this pixel/prototype with a set of reliable reference points. In essence, the refer- ential correlation reflects the consensus among reliable ref-erence points with a broader receptive field and thus it en- codes the relative semantic comparability of the reference points that can be relied upon, which is from a different per- spective than the absolute pairwise prototype-pixel correla- tion. Intuitively, each pair of true prototype-pixel correla- tion ( e.g., thef1-p1pair in Fig. 1 (c) derived from the proto- types and mitochondria images should be not only visually similar to each other ( i.e., high direct pairwise prototype- pixel correlation), but also similar to any other reference point ( i.e., similar referential correlation pair). Moreover, we assemble referential correlation into the cross-attention mechanism with the ability to capture long-range dependen- cies. In this case, the relatively equivocal pixels ( e.g., the f1-p2pair in Fig. 1 (c) will be suppressed while the reliable ones are highlighted to reduce the correspondence noise. In the reliable prototype selection module (RPrS), in order to further evaluate the reliability of prototypes in construct- ing prototype-level consistency regularization, we draw in- spiration from bayesian deep learning [12] and devise a reliability-aware consistency loss to pursue implicitly learn the reliability about each prototype in a data-driven way. In this way, the equivocal prototypes will be suppressed while the reliable ones are highlighted in the supervision signals. In this work, our contributions can be concluded as fol- lows: (1) To the best of our knowledge, this is the first work to rethink how to achieve effective consistency regulariza- tion for semi-supervised mitochondria segmentation, from the perspective of reliable prototype-level supervision. We analyze the gap between mitochondrial images and natu- ral images, hoping our work will provide some insight for researchers in this field. (2) We propose a dual-reliable (DualRel) network in a unified framework. Specifically, we design the reliable pixel aggregation module to rectify the direct pairwise correlation, the reliable prototype se- lection module to further evaluate the reliability of proto- types in constructing prototype-level consistency regular- ization. (3) Extensive experimental results on three chal- lenging benchmarks demonstrate that our method performs favorably against state-of-the-art semi-supervised segmen- tation methods. Importantly, with extremely few samples used for training, DualRel is also on par with current state- of-the-art fully supervised methods.

Ma_Symmetric_Shape-Preserving_Autoencoder_for_Unsupervised_Real_Scene_Point_Cloud_Completion_CVPR_2023

Abstract Unsupervised completion of real scene objects is of vi- tal importance but still remains extremely challenging in preserving input shapes, predicting accurate results, and adapting to multi-category data. To solve these prob- lems, we propose in this paper an Unsupervised Symmetric Shape-Preserving Autoencoding Network, termed USSPA, to predict complete point clouds of objects from real scenes. One of our main observations is that many natural and man-made objects exhibit significant symmetries. To ac- commodate this, we devise a symmetry learning module to learn from those objects and to preserve structural sym- metries. Starting from an initial coarse predictor, our au- toencoder refines the complete shape with a carefully de- signed upsampling refinement module. Besides the discrim- inative process on the latent space, the discriminators of our USSPA also take predicted point clouds as direct guid- ance, enabling more detailed shape prediction. Clearly different from previous methods which train each category separately, our USSPA can be adapted to the training of multi-category data in one pass through a classifier-guided discriminator, with consistent performance on single cate- gory. For more accurate evaluation, we contribute to the community a real scene dataset with paired CAD models as ground truth. Extensive experiments and comparisons demonstrate our superiority and generalization and show that our method achieves state-of-the-art performance on unsupervised completion of real scene objects.

1. Introduction As the standard outputs of 3D scanners [12, 32], point clouds are becoming more and more popular [9] which are also the basic data structure in 3D geometry process- ing [4, 5,13]. Complete point clouds are hard to obtain due to the nature of the scanning process and object oc- clusion [35]. Due to the defects of incomplete point clouds on downstream applications such as reconstruction [10], re- cent works [17, 19,22,23,26,30,33,35] pay more attention *Corresponding author. (b) (c)Input Unpaired Ours (a)Disp3D Completion ResultsReal Scene ObjectsShapeInv Figure 1. Visual comparison of predicted results on real scene data by our USSPA and other works (top) and our complete result on a whole real scene (bottom). (a) shows an example of a real scene partial point cloud of a chair and the complete predictions by Disp3D [23], ShapeInv [36], Unpaired [32] and our method. As shown, our prediction result is more accurate and uniform accord- ing to the input, which contains complete arms and legs. (b) and (c) show the original point cloud of a real scene and the complete results of all the objects in this scene. to point cloud completion which relies on paired artificial complete point clouds for supervised training to complete partial point clouds. However, these supervised works are difficult to apply in practice because of the great gap be- tween artificial data and real scene data and the inaccessi- bility to the ground truth of real scene data. Therefore, it is important to complete partial point clouds from real scene in an unsupervised way. Recent unsupervised works [24, 32, 36] only require real 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. 13560 scene partial point clouds and artificial CAD models for un- paired completion utilizing GANs [8] as their fundamental frameworks, most of which need pre-training on artificial data. The main ideas are to transform latent codes from the space of real scene partial data to the space of artificial complete data and then employ the decoder trained on arti- ficial data to predict the complete point cloud. Essentially, these methods make the predictions whose distributions are consistent with the artificial models. Most of them, how- ever, just extract a global feature from the partial input with- out fully exploiting its geometry information, leading to the prediction severely deviated from the input. And such in- formation actually provides vital clues and constraints for completion. Furthermore, prediction results by these meth- ods usually lack enough geometric details due to the ab- sence of an explicit discriminative process on point clouds. These domain transforming methods are also hard to adapt to multi-category data or other datasets. In this paper, we present an unsupervised symmetric shape-preserving autoencoding network, termed USSPA, for the completion of real scene objects, as shown in Figure 2, which is a GAN-based end-to-end network without the requirement of pre-trained weights. Different from previ- ous domain transforming methods which cannot fully lever- age existing incomplete models, we argue that the exist- ing partial scanning, which also provides vital clues and constraints for the prediction of the missing part, should be preserved to some extent. To this end, we exploit the symmetries shown in many natural or man-made objects and devise a novel symmetry learning module to generate symmetrical point clouds of existing parts by predicting the symmetric planes. This enables our network to preserve the shapes of input symmetrically, intrinsically facilitating structure completion, as shown in Figure 1. For those parts that can not be directly inferred from inputs, we employ an initial coarse module for an initial prediction first. Start- ing from the initial guess, we specifically design a refine- ment autoencoder with an upsampling refinement module for detailed refinement and the local feature grouping for extracting local information, to learn detailed structures of artificial data through the autoencoding process. Benefit- ing from this, our final prediction is accurate, uniform, and symmetric shape-preserving. Besides the indirect guidance of the feature discriminator on latent space, our point dis- criminator takes predicted point clouds as direct guidance for generating more accurate shapes. Compared with pre- vious methods which train each category separately, our method can classify the partial point clouds simultaneously through a classifier-guided discriminator when adapted to multi-category data, with consistent performance on the sin- gle category. To measure the performance of unsupervised comple- tion quantitatively, we build a dataset from ScanNet [5] andShapeNet [2] utilizing the annotations of Scan2CAD [1]. Our dataset contains real scene partial point clouds and paired ground truths that are only used for evaluation in our experiments. Extensive comparisons against previous works on this dataset and the public PCN Dataset [35] show the superiority and generalization of our method which achieves state-of-the-art performance on unsupervised com- pletion of real scene objects. Our main contributions are as follows. • We propose a novel USSPA for unsupervised real scene point cloud completion whose prediction is accurate, uniform and symmetric shape-preserving. Clearly different from previous works training each category separately, our USSPA can be adapted to the training of multi-category data in one pass by classify- ing the input simultaneously. • We propose a novel symmetry learning module and a novel refinement autoencoder. The symmetry learn- ing module preserves input shapes by generating sym- metrical point clouds, and the refinement autoencoder learns the detailed information from artificial data to refine the initial guess by an autoencoding process. • We propose a new evaluation method for obtaining paired ground truths and partial data from artificial and real scene datasets using alignment information, which can be used to more accurately evaluate unsupervised completion of real scene objects.

Pan_Deep_Discriminative_Spatial_and_Temporal_Network_for_Efficient_Video_Deblurring_CVPR_2023

Abstract How to effectively explore spatial and temporal informa- tion is important for video deblurring. In contrast to exist- ing methods that directly align adjacent frames without dis- crimination, we develop a deep discriminative spatial and temporal network to facilitate the spatial and temporal fea- ture exploration for better video deblurring. We first de- velop a channel-wise gated dynamic network to adaptively explore the spatial information. As adjacent frames usually contain different contents, directly stacking features of ad- jacent frames without discrimination may affect the laten- t clear frame restoration. Therefore, we develop a simple yet effective discriminative temporal feature fusion module to obtain useful temporal features for latent frame restora- tion. Moreover, to utilize the information from long-range frames, we develop a wavelet-based feature propagation method that takes the discriminative temporal feature fu- sion module as the basic unit to effectively propagate main structures from long-range frames for better video deblur- ring. We show that the proposed method does not require additional alignment methods and performs favorably a- gainst state-of-the-art ones on benchmark datasets in terms of accuracy and model complexity.

1. Introduction With the rapid development of hand-held video captur- ing devices in our daily life, capturing high-quality clear videos becomes more and more important. However, due to the moving objects, camera shake, and depth variation dur- ing the exposure time, the captured videos usually contain significant blur effects. Thus, there is a great need to restore clear videos from blurred ones so that they can be pleasantly viewed on display devices and facilitate the following video understanding problems. Different from single image deblurring that explores spa- Co-first authorship yCorresponding author 2200 2400 ···'671HW2XUV &'9'763 67)$1'9'6() (67511)*67 ('955HVWRUPHU %DVLF965 0351HWFigure 1. Floating point operations (FLOPs) vs. video deblurring performance on the GoPro dataset [24]. Our model achieves fa- vorable results in terms of accuracy and FLOPs. tial information for blur removal, video deblurring is more challenging as it needs to model both spatial and tempo- ral information. Conventional methods usually use optical flow [2,9,14,36] to model the blur in videos and then joint- ly estimate optical flow and latent frames under the con- straints by some assumed priors. As pointed out by [26], these methods usually lead to complex optimization prob- lems that are difficult to solve. In addition, improper priors will significantly affect the quality of restored videos. Instead of using assumed priors, lots of methods develop kinds of deep convolutional neural networks (CNNs) to ex- plore spatial and temporal information for video deblurring. Several approaches stack adjacent frames as the input of C- NN models [29] or employ spatial and temporal 3D convo- lution [40] for latent frame restoration. Gast et al. [10] show that using proper alignment strategies in deep CNNs would improve deblurring performance. To this end, several meth- ods introduce alignment modules in deep neural networks. The commonly used alignment modules for video deblur- ring mainly include optical flow [26], deformable convolu- tion [32], and so on. However, estimating alignment infor- mation from blurred adjacent frames is not a trivial task due to the influence of motion blur. In addition, using align- ment modules usually leads to large deep CNN models 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. 22191 are difficult to train and computationally expensive. For ex- ample, the CDVDTSP method [26] with optical flow as the alignment module has 16.2 million parameters with FLOP- s of 357.79G while the EDVR method [32] using the de- formable convolution as the alignment module has 23.6 mil- lion parameters with FLOPs of 2298.97G. Therefore, it is of great interest to develop a lightweight deep CNN model with lower computational costs to overcome the limitation- s of existing alignment methods in video deblurring while achieving better performance. Note that most existing methods restore each clear frame based on limited local frames, where the temporal infor- mation from non-local frames is not fully explored. To overcome this problem, several methods employ recurrent neural networks to better model temporal information for video deblurring [15]. However, these methods have limit- ed capacity to transfer the useful information temporally for latent frame restoration as demonstrated in [41]. To rem- edy this limitation, several methods recurrently propagate information of non-local frames with some proper atten- tion mechanisms [41]. However, if the features of non-local frames are not estimated correctly, the errors will accumu- late in the recurrent propagation process, which thus affects video deblurring. As the temporal information exploration is critical for video deblurring, it is a great need to devel- op an effective propagation method that can discriminative- ly propagate useful information from non-local frames for better video restoration. In this paper, we develop an effective deep discriminative spatial and temporal network (DSTNet) to distinctively ex- plore useful spatial and temporal information from videos for video deblurring. Motivated by the success of multi- layer perceptron (MLP) models that are able to model glob- al contexts, we first develop a channel-wise gated dynamic network to effectively explore spatial information. In addi- tion, to exploit the temporal information, instead of directly stacking estimated features from adjacent frames without discrimination, we develop a simple yet effective discrim- inative temporal feature fusion module to fuse the features generated by the channel-wise gated dynamic network so that more useful temporal features can be adaptively ex- plored for video deblurring. However, the proposed discriminative temporal feature fusion module does not utilize the information from long- range frames. Directly repeating this strategy in a recur- rent manner is computationally expensive and may propa- gate and accumulate the estimation errors of features from long-range frames, leading to adverse effects on the final video deblurring. To solve this problem, we develop a wavelet-based feature propagation method that effective- ly propagates main structures from long-range frames for better video deblurring. Furthermore, the deep discrimina- tive spatial and temporal network does not require addition-al alignment modules (e.g., optical flow used in [26], de- formable convolution used in [32]) and is thus efficient yet effective for video deblurring as shown in Figure 1. The main contributions are summarized as follows: We propose a channel-wise gated dynamic network (CWGDN) based on multi-layer perceptron (MLP) models to explore the spatial information. A detailed analysis demonstrates that the proposed CWGDN is more effective for video deblurring. We develop a simple yet effective discriminative tem- poral feature fusion (DTFF) module to explore useful temporal features for clear frame reconstruction. We develop a wavelet-based feature propagation (WaveletFP) method to efficiently propagate useful structures from long-range frames and avoid error ac- cumulation for better video deblurring. We formulate the proposed network in an end-to-end trainable framework and show that it performs favor- ably against state-of-the-art methods in terms of accu- racy and model complexity.

Qiao_End-to-End_Vectorized_HD-Map_Construction_With_Piecewise_Bezier_Curve_CVPR_2023

Abstract Vectorized high-definition map (HD-map) construction, which focuses on the perception of centimeter-level environ- mental information, has attracted significant research inter- est in the autonomous driving community. Most existing ap- proaches first obtain rasterized map with the segmentation- based pipeline and then conduct heavy post-processing for downstream-friendly vectorization. In this paper, by delving into parameterization-based methods, we pioneer a concise and elegant scheme that adopts unified piecewise B ´ezier curve. In order to vectorize changeful map elements end- to-end, we elaborate a simple yet effective architecture, named Piecewise B ´ezier HD-map Network ( BeMapNet ), which is formulated as a direct set prediction paradigm and postprocessing-free. Concretely, we first introduce a novel IPM-PE Align module to inject 3D geometry prior into BEV features through common position encoding in Transformer. Then a well-designed Piecewise B ´ezier Head is proposed to output the details of each map element, including the coor- dinate of control points and the segment number of curves. In addition, based on the progressively restoration of B ´ezier curve, we also present an efficient Point-Curve-Region Loss for supervising more robust and precise HD-map modeling. Extensive comparisons show that our method is remarkably superior to other existing SOTAs by 18.0mAP at least1.

1. Introduction As one of the most fundamental components in the auto- driving system, high-definition map contains centimeter de- tails of traffic elements, vectorized topology and navigation information, which instruct ego-vehicle to accurately locate itself on the road and understand what is coming up ahead. At present, conventional SLAM-based solutions [45, 46, 60] have been widely adopted in practice. Yet, due to dilemmas of high annotation costs and untimely updates, the offline approach is gradually being replaced by the learning-based online HD-map construction with onboard sensors. *Corresponding author . 1https://github.com/er-muyue/BeMapNet DividerCrossingBoundary(a) HD-map example (c) map curve restored by real control points (b) vanillaBéziervs. piecewiseBézier Figure 1. Illustration of our motivation for piecewise B ´ezier curve, termed as ⟨k,n ⟩, where kis the piece number and nis the degree. Fig.(a) is a real HD-map case from NuScenes .Fig.(b) compares the difference between vanilla and piecewise B ´ezier curve through the same map element, where the light purple is the restored curve with B ´ezier process. The last is more efficient than previous ones with reducing the number of control points by 64%in this case. Fig.(c) illustrates that piecewise B ´ezier curve can model arbitrary- shaped curves. Note the blue circles denote actual control points. The deep-based paradigm of online HD-map building is gradually developing, but it still faces two main challenges: 1)modeling instance-level vectorized HD-map end-to-end. Most existing works construct HD-map by rasterizing BEV (bird-eye-view) maps into semantic pixels with segmenta- tion [24,42], which not only lacks the modeling of instance- level details, but also requires heavy post-processing to ob- tain vectorized information. As a sub-task, lane detection makes a relatively better advance for this issue, that is, in ad- dition to segmentation-based methods [39,41,62], there are also point-based [25,47] and curve-based [12,28] schemes. However, compared to the simple lane scenario, HD-map contains more shape-changeful elements, so such methods cannot be directly adopted into the HD-map construction. 2)performing 2D-3D perspective transformation efficiently. Obtaining 3D-BEV perception from multi-view 2Dimages is an essential step for building HD-map , which is mainly divided into three ideas, i.e. geometric priors [44], learnable parameters [40, 42], and a combination of the two [17, 43]. Note the assumptions of geometry-based methods often 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. 13218 not conform to the actual situation, leading to such schemes are less adaptable, while learning-based methods require a large amount of labeled data to generalize across various scenarios. Combining the above two branches not only has multi-scenario scalability, but also reduces the demand for annotated data, has attracted increasing research interest. To the best of our knowledge, the curve parameterization construction of HD-map in the BEV space is vacancy and no one has explored it. Based on the widely used B ´ezier curve, which is mathematically defined by a set of control points, we pioneer to devise a concise and elegant HD-map scheme that adopts piecewise B´ezier curve, where each map curve is divided into multiple ksegments and each segment is then represented by a vanilla B ´ezier curve with degree n, hence denoted as ⟨k,n ⟩. Despite ⟨1,n⟩is enough to express any map element with infinite nin theory, more complex curve tends to require higher degree, meaning that there are more control points need to be modeled, which is shown in Fig.1. The proposed piecewise strategy allows us to parameterize a curve more compactly with fewer control points and higher capacity, which is extremely scalable and robust in practice. Inspired by the above motivations, we propose an end-to- end vectorized HD-map construction architecture, named asPiecewise B´ezierHD-map Network ( BeMapNet ). The overall framework is illustrated in detail in Fig.2, which streamlines the architecture into four primary modules for gradually-enriched information, i.e. feature extractor shared among multi-view images, semantic BEV decoder for 2D- 3Dperspective elevation, instance B ´ezier decoder for curve- level descriptors, and piecewise B ´ezier head for point-level parameterization. To be concrete, we first introduce a novel IPM-PE Align module into Transformer-based decoders, which injects IPM (inverse perspective mapping) geomet- ric priors into BEV features via PE(position encoding) and hardly adds any parameters except a FClayer. Secondly, we further design a Piecewise B ´ezier Head for dynamic curve modeling with adopting two branches as classification and regression, where the former classifies the number of piece to determine the curve length and the latter regresses the coordinate of control points to determine the curve shape. Lastly, we present an Point-Curve-Region Loss for robust curve modeling by supervising restoration information as a progressive manner. Since it is modeled as a sparse set pre- diction task and optimized with a bipartite matching loss, our method is postprocessing-free and high-performance. The main contributions of our approach are three-folds: • We pioneer the BeMapNet for concise and elegant mod- eling of HD-map with unified piecewise B ´ezier curve. • We elaborate the overall end-to-end architecture with in- novatively introducing IPM-PE Align Module ,Piecewise B´ezier Output Head and well-designed PCR-Loss . •BeMapNet is remarkably superior to SOTAs on existing benchmarks, revealing the effectiveness of our approach.

Peng_Hierarchical_Dense_Correlation_Distillation_for_Few-Shot_Segmentation_CVPR_2023

Abstract Few-shot semantic segmentation (FSS) aims to form class-agnostic models segmenting unseen classes with only a handful of annotations. Previous methods limited to the semantic feature and prototype representation suffer from coarse segmentation granularity and train-set overfitting. In this work, we design Hierarchically Decoupled Match- ing Network (HDMNet) mining pixel-level support corre- lation based on the transformer architecture. The self- attention modules are used to assist in establishing hierar- chical dense features, as a means to accomplish the cascade matching between query and support features. Moreover, we propose a matching module to reduce train-set over- fitting and introduce correlation distillation leveraging se- mantic correspondence from coarse resolution to boost fine- grained segmentation. Our method performs decently in ex- periments. We achieve 50.0%mIoU on COCO- 20idataset one-shot setting and 56.0%on five-shot segmentation, re- spectively. The code is available on the project website1.

1. Introduction Semantic segmentation tasks [2, 3, 22, 52] have made tremendous progress in recent years, benefiting from the rapid development of deep learning [13,32]. However, most existing deep networks are not scalable to previously unseen classes and rely on annotated datasets to achieve satisfy- ing performance. Data collection and annotation cost much time and resources, especially for dense prediction tasks. Few-shot learning [34, 39, 43] has been introduced into semantic segmentation [5, 38] to build class-agnostic mod- els quickly adapting to novel classes. Typically, few- shot segmentation (FSS) divides the input into the query and support sets [5, 46, 48, 51] following the episode paradigm [41]. It segments the query targets conditioned on *Corresponding Author 1https://github.com/Pbihao/HDMNet PASCAL-𝟓𝒊COCO-𝟐𝟎𝒊SupportQueryOursBaselineFigure 1. Activation maps of the correlation values on both PASCAL- 5i[29] and COCO- 20i[26]. The baseline is prone to give high activation values to the categories sufficiently witnessed during training, such as the “People” class, even with other sup- port annotations. Then we convert it to the hierarchically decou- pled matching structure and adopt correlation map distillation to mine inner-class correlation. the semantic clues from the support annotations with meta- learning [34, 39] or feature matching [25, 41, 49]. Previous few-shot learning methods may still suffer from coarse segmentation granularity and train-set over- fitting [38] issues. As shown in Fig. 1, “people” is the base class that has been sufficiently witnessed during train- ing. But the model is still prone to yield high activa- tion to “people” instead of more related novel classes with the support samples, producing inferior results. This is- sue stems from framework design, as illustrated in Fig. 2. Concretely, prototype-based [38,42] and adaptive-classifier methods [1, 23] aim at distinguishing different categories with global class-wise characteristics. It is challenging to compute the correspondence of different components be- tween query and support objects for the dense prediction tasks. In contrast, matching-based methods [49] mine pixel- level correlation but may heavily rely on class-specific fea- 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. 23641 𝐼!𝐼"EncoderSelf-AttnCross-Attn𝑀"×𝑁Encoder𝐼!PoolExpandConcatSelf-AttnSelf-AttnSelf-AttnCorrelation&Distillation(a)(c)(d)𝐼"𝑀"Encoder𝐼"𝑀"𝐼!Encoder(b)DecoderAdaptiveClassifier𝐼!𝐼"𝑀"Figure 2. Illustration of different few-shot segmentation frame- works. (a) Prototype-based method. (b) Adaptive-classifier method. (c) Feature matching with transformer architecture. (d) Our Hierarchically Decoupled Matching Network (HDMNet) with correlation map distillation. tures and cause overfitting and weak generalization. To address these issues, we propose Hierarchically De- coupled Matching Network (HDMNet) with correlation map distillation for better mining pixel-level support corre- spondences. HDMNet extends transformer architecture [6, 40,44] to construct the feature pyramid and performs dense matching. Previous transformer-based methods [35, 49] adopt the self-attention layer to parse features and then feed query and support features to the cross-attention layer for pattern matching, as illustrated in Fig. 2(c). This pro- cess stacks the self- and cross-attention layers multiple times, mixes separated embedding features, and acciden- tally causes unnecessary information interference. In this paper, we decouple the feature parsing and match- ing process in a hierarchical paradigm and design a new matching module based on correlation and distillation. This correlation mechanism calculates pixel-level corre- spondence without directly relying on the semantic-specific features, alleviating the train-set overfitting problem. Fur- ther, we introduce correlation map distillation [14, 50] that encourages the shallow layers to approximate the semantic correlation of deeper layers to make the former more aware of the context for high-quality prediction. Our contribution is the following. 1) We extend the transformer to hierarchical parsing and feature matching for few-shot semantic segmentation, with a new matching mod- ule reducing overfitting. 2) We propose correlation map dis- tillation leveraging soft correspondence under multi-level and multi-scale structure. 3) We achieve new state-of- the-art results on standard benchmark of COCO- 20iand PASCAL- 5iwithout compromising efficiency.

Ma_CREPE_Can_Vision-Language_Foundation_Models_Reason_Compositionally_CVPR_2023

Abstract A fundamental characteristic common to both human vi- sion and natural language is their compositional nature. Yet, despite the performance gains contributed by large vision and language pretraining, we find that—across 7 architec- tures trained with 4 algorithms on massive datasets—they struggle at compositionality. To arrive at this conclusion, we introduce a new compositionality evaluation benchmark, CREPE, which measures two important aspects of compo- sitionality identified by cognitive science literature: system- aticity and productivity. To measure systematicity, CREPE consists of a test dataset containing over 370Kimage-text pairs and three different seen-unseen splits. The three splits are designed to test models trained on three popular training datasets: CC-12M, YFCC-15M, and LAION-400M. We also generate 325K,316K, and 309Khard negative captions for a subset of the pairs. To test productivity, CREPE con- tains 17Kimage-text pairs with nine different complexities plus278Khard negative captions with atomic, swapping and negation foils. The datasets are generated by repurpos- ing the Visual Genome scene graphs and region descriptions and applying handcrafted templates and GPT-3. For sys- tematicity, we find that model performance decreases con- sistently when novel compositions dominate the retrieval set, with Recall@1 dropping by up to 9%. For productivity, models’ retrieval success decays as complexity increases, frequently nearing random chance at high complexity. These results hold regardless of model and training dataset size.

1. Introduction Compositionality, the understanding that “the meaning of the whole is a function of the meanings of its parts” [11], is held to be a key characteristic of human intelligence. In language, the whole is a sentence, made up of words. In vision, the whole is a scene, made up of parts like objects, their attributes, and their relationships [31, 35]. *Equal contribution Figure 1. We introduce CREPE, a benchmark to evaluate whether vision-language foundation models demonstrate two fundamental aspects of compositionality: systematicity and productivity. To eval- uate systematicity, CREPE utilizes Visual Genome and introduces three new test datasets for the three popular pretraining datasets: CC-12M, YFCC-15M, and LAION-400M. These enable evaluating models’ abilities to systematically generalize their understanding to seen compounds, unseen compounds, and even unseen atoms. To evaluate productivity, CREPE introduces examples of nine com- plexities, with three types of hard negatives for each. Through compositional reasoning, humans can understand new scenes and generate complex sentences by combining known parts [6, 27, 30]. Despite compositionality’s impor- tance, there are no large-scale benchmarks directly evaluat- ing whether vision-language models can reason composition- ally. These models are pretrained using large-scale image- caption datasets [62, 64, 74], and are already widely applied for tasks that benefit from compositional reasoning, includ- ing retrieval, text-to-image generation, and open-vocabulary classification [10,57,60]. Especially as such models become ubiquitous “foundations” for other models [5], it is critical to understand their compositional abilities. Previous work has evaluated these models using image- text retrieval [32,56,82]. However, the retrieval datasets used either do not provide controlled sets of negatives [45, 74] or study narrow negatives which vary along a single axis 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. 10910 Figure 2. An overview of the systematicity retrieval set generation process. First, a model’s image-caption training set is parsed to identify what atoms and compounds the model has seen. Then, an evaluation set is divided into three compositional splits according to whether the model has seen all the compounds (Seen Compounds), only all the atoms of the caption (Unseen Compounds), or neither (Unseen Atoms). Finally, hard negative captions HN-A TOM and HN-C OMP are generated for the hard negatives retrieval set DHN test. (e.g. permuted word orders or single word substitutions as negative captions) [21, 51, 65, 75]. Further, these analyses have also not studied how retrieval performance varies when generalizing to unseen compositional combinations, or to combinations of increased complexity. We introduce CREPE (Compositional REPresentation Evaluation): a new large-scale benchmark to evaluate two aspects of compositionality: systematicity andproductivity (Figure 1). Systematicity measures how well a model is able to represent seen versus unseen atoms and their composi- tions. Productivity studies how well a model can compre- hend an unbounded set of increasingly complex expressions. CREPE uses Visual Genome’s scene graph representation as the compositionality language [35] and constructs evaluation datasets using its annotations. To test systematicity, we parse the captions in three popular training datasets, CC-12M [8], YFCC-15M [74], and LAION-400M [62], to identify atoms (objects, relations, or attributes) and compounds (combina- tions of atoms) present in each dataset. For each training set, we curate corresponding test sets containing 385K,385K and373Kimage-text pairs respectively, with splits checking generalization to seen compounds, unseen compounds, and unseen atoms. To test productivity, CREPE contains 17K image-text pairs split across nine levels of complexity, as defined by the number of atoms present in the text. Exam- ples across all datasets are paired with various hard negative types to ensure the legitimacy of our conclusions. Our experiments—across 7 architectures trained with 4 training algorithms on massive datasets—find that vision- language models struggle at compositionality, with both systematicity and productivity. We present six key findings: first, our systematicity experiments find that models’ perfor- mance consistently drops between seen and unseen composi- tions; second, we observe larger drops for models trained onLAION-400M (up to a 9%decrease in Recall@1); third, our productivity experiments indicate that retrieval performance degrades with increased caption complexity; fourth, we find no clear trend relating training dataset size to models’ com- positional reasoning; fifth, model size also has no impact; finally, models’ zero-shot ImageNet classification accuracy correlates only with their absolute retrieval performance on the systematicity dataset but not systematic generalization to unseen compounds or to productivity.1

Nagata_Tangentially_Elongated_Gaussian_Belief_Propagation_for_Event-Based_Incremental_Optical_Flow_CVPR_2023

Abstract Optical flow estimation is a fundamental functionality in computer vision. An event-based camera, which asyn- chronously detects sparse intensity changes, is an ideal device for realizing low-latency estimation of the optical flow owing to its low-latency sensing mechanism. An existing method using local plane fitting of events could utilize the sparsity to realize incremental updates for low-latency estimation; however, its output is merely a normal component of the full optical flow. An alterna- tive approach using a frame-based deep neural network could estimate the full flow; however, its intensive non- incremental dense operation prohibits the low-latency estimation. We propose tangentially elongated Gaussian (TEG) belief propagation (BP) that realizes incremental full-flow estimation. We model the probability of full flow as the joint distribution of TEGs from the normal flow measurements, such that the marginal of this distribution with correct prior equals the full flow. We formulate the marginalization using a message-passing based on the BP to realize efficient incremental updates using sparse measurements. In addition to the theoretical justification, we evaluate the effectiveness of the TEGBP in real-world datasets; it outperforms SOTA incremental quasi-full flow method by a large margin. (The code is available at https://github.com/DensoITLab/tegbp/ ).

1. Introduction Optical flow estimation, which computes the correspon- dence of pixels in different time measurements, is a fun- damental building block of computer vision. One needs to estimate the flow at low latency in many practical appli- cations, such as autonomous driving cars, unmanned aerial vehicles, and factory automation robots. Most of the ex- isting optical flow algorithm utilizes dense video frames; it computes the flow by searching the similar intensity pat- tern [ 15,29]. Recently, methods using deep neural net- work (DNN) [ 29] demonstrate impressive accuracy at the †These authors contributed equally to this work. Figure 1. Overview of the proposed TEGBP . We model a belief about the full flow from a normal flow measurement using TEG (RGBellipse). The mean of the marginal distribution of each TEG with an appropriate prior equals the full flow ( magenta arrow). The marginal ( magenta ellipse) is computed incrementally using local message passing ( RGBarrows) based on BP. cost of higher-computational cost. Either model-based or DNN-based, the frame-based algorithm needs to compute the entire pixel for every frame, even when there are subtle changes or no changes at all. This dense operation makes it difficult to realize low-latency estimation, especially on the resource-constrained edge device. The event-based camera is a bio-inspired vision sen- sor, which asynchronously detects intensity change on each pixel. Thanks to the novel sensing mechanism, the camera equips favorable characteristics for optical flow estimation, such as high dynamic range (HDR), blur-free measurement, and, most importantly, sparse low-latency data acquisition. Many researchers have explored the way to utilize sparsity to realize efficient low-latency estimation; one extends the well-known Lucas-Kanade algorithm [ 7], and the other ex- ploits the local planer shape of the spatiotemporal event streams [ 6]. These methods could utilize the sparsity for efficient incremental processing; however, the optical flow computed in this way (e.g., by plane fitting) is the normal flow, which is a normal component of full flow and often different from them1we want to obtain. The normal flow is the component of the full flow perpendicular to the edge (i.e., parallel to the intensity gradient). Some work tried to recover the full flow from the normal flow [ 2]; however, it does not precisely equal the full flow (refer to Sec. 2). There exist methods that could estimate full flow, such as a varia- 1full flow is usually simply called optical flow orflow, yet, we use full flowwhen we want to highlight the difference with the normal flow . 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. 21940 tional method [ 5], a multi-scale extension of contrast max- imization [ 25], or a frame-based DNN [ 14]. However, they need to apply non-incremental dense operation for all the pixels of the event frame (dense representation constructed from sparse events) for every frame, which prohibits the low-latency estimation on the edge device. Our research goal is to realize an incremental full flow algorithm from sparse normal flow measurements. To this end, we propose Tangentially Elongated Gaussian (TEG) Belief Propagation (BP) . We compute the full flow using the normal flow measurements, which can be observed di- rectly from an event camera or computed cheaply using an existing algorithm2. Notice that given a single measurement of normal flow, there are infinite possibilities for the full flow along the tangential direction of the normal flow. We model this uncertainty using the TEG, Gaussian distribu- tion, which has a large variance along the tangential direc- tion of the normal flow. The probability density of full flow on each pixel is given as the marginals of the joint distribu- tion of TEG data factor and some prior factor on a sparse graph (Fig. 1, Sec. 3.3.2 ). We leverage the sparse graph to formulate the incremental full flow estimation algorithm using message-passing based on BP [ 9]. We evaluate the ef- fectiveness of the TEGBP on real-world data captured from aerial drones and automobiles. TEGBP outperforms SOTA incremental method [ 2] by a large margin.

Nguyen_TIPI_Test_Time_Adaptation_With_Transformation_Invariance_CVPR_2023

Abstract When deploying a machine learning model to a new en- vironment, we often encounter the distribution shift prob- lem – meaning the target data distribution is different from the model’s training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distri- bution can affect the model’s performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a sur- rogate loss on the test-time unlabeled target data. In par- ticular, minimizing the prediction’s entropy on target sam- ples [34] has received much interest as it is task-agnostic and does not require altering the model’s training phase (e.g., does not require adding a self-supervised task dur- ing training on the source domain). However, as the tar- get data’s batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not op- timal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an in- variance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model’s performance under input transformations. The resulting method (TIPI – Test tIme adaPtation with transfor- mation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet- C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI. *The last two authors contributed equally

1. Introduction Distribution shift is a common problem and is often faced in real-world applications. Specifically, despite tak- ing various precautions while training a machine learning model to ensure a better generalization (e.g., collecting and training on multiple source domains [17, 27], finding flat minima [5], training with meta-learning objectives [16] etc.), the model often still struggles when the test data dis- tribution shifts slightly. Note that it is also common not to have labels for the new shifted domain during test time. To tackle this problem, test time adaptation (also known as online domain adaptation) is a framework that allows the model to adapt to the target distribution using unlabeled test data batches. This is necessary since we typically do not have time to annotate the data during test time and can only make use of the unlabeled data. In this framework, the model needs to give predictions for the target data while simultaneously updating itself to improve its performance on that particular target distribution. We assume a situation in which the target data only arrive in small batches, which makes the adaptation task extremely challenging and ren- ders traditional domain adaptation techniques such as rep- resentation alignment (via a distance metric) ineffective. Within this test time adaptation framework, a common and effective approach is to optimize a surrogate objective function in lieu of the true loss function on the target data. The first group of surrogate objectives is the loss functions of self-supervised tasks. In particular, one would formu- late a user-defined task (such as predicting the rotation an- gle of an image) and train it alongside the main task on the source domain; and keep training the self-supervised task on the test-time target data [19, 33]. However, these are not fully test-time adaptation methods, since they re- quire altering the training procedure of the source domain. The second line of surrogate objectives is unsupervised loss functions. Among this group of unsupervised objectives, entropy minimization (TENT) [34] is the most successful method, and has been shown to be consistent across many benchmarks. Furthermore, different from the former 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. 24162 TENT is a fully test-time adaptation method. For these rea- sons, TENT has received much interest and a lot of follow- up papers/discussions. However, as pointed out by its authors, TENT is not ro- bust when using small batch sizes, as it often collapses to a trivial solution (i.e., it always predicts the same class for all input). This is detrimental to real-world applications since test data often arrive in small batches. For example, au- tonomous driving systems only process a few frames in real time – they typically do not accumulate the frames (let’s say within one minute) to form a bigger batch. Note that TENT is previously evaluated mainly for large batch sizes (such as 64,128 or 200). In this paper, we aim to tackle the aforementioned prob- lem. Specifically, we aim to develop an unsupervised sur- rogate objective function for the test time adaptation prob- lem such that it is task-agnostic, does not require altering the training procedure (e.g., does not require incorporating a self-supervised task into the training process), and is more resilient against small batch sizes. We first provide theoretical results regarding a model’s performance under input transformations. Specifically, we show that a model’s loss on a data distribution is bounded by the KL distance on the predictive distribution of the data before and after the transformations (which we will use as a regularizer), and its loss on the transformed data distri- bution. Motivated by this result, we use small shifts in the input images that can simulate real source-target shifts, and enforce the network to be invariant under such data transfor- mations. Our model outperforms TENT (and other relevant baselines) in all problem settings considered in the paper and is remarkably robust in the small-batch-size regime. Our contributions in this paper are threefold: • We provide theoretical results regarding a model’s performance under transformations of the input data. Specifically, a model’s performance on the target do- main is bounded by an invariance term (maximum KL divergence of the predictive distributions before and after the transformations) and its loss on the trans- formed domain. • We propose to find input transformations that can sim- ulate the domain shifts, and enforce the network to be invariant under such transformations, using a reg- ularizer based on our derived bound. The resulting method is TIPI (Test tIme ada Ptation with transfor- mation Invariance). • We perform extensive experiments on a wide range of datasets (Cifar10-C, Cifar100-C, ImageNet-C, DIG- ITS, and VisDA17) and settings (varying batch sizes) to validate our method. TIPI shows preferable perfor- mance compared to relevant baselines.

Li_Rethinking_Out-of-Distribution_OOD_Detection_Masked_Image_Modeling_Is_All_You_CVPR_2023

Abstract The core of out-of-distribution (OOD) detection is to learn the in-distribution (ID) representation, which is dis- tinguishable from OOD samples. Previous work applied recognition-based methods to learn the ID features, which tend to learn shortcuts instead of comprehensive repre- sentations. In this work, we find surprisingly that simply using reconstruction-based methods could boost the per- formance of OOD detection significantly. We deeply ex- plore the main contributors of OOD detection and find that reconstruction-based pretext tasks have the potential to pro- vide a generally applicable and efficacious prior, which benefits the model in learning intrinsic data distributions of the ID dataset. Specifically, we take Masked Image Mod- eling as a pretext task for our OOD detection framework (MOOD). Without bells and whistles, MOOD outperforms previous SOTA of one-class OOD detection by 5.7%, multi- class OOD detection by 3.0%, and near-distribution OOD detection by 2.1%. It even defeats the 10-shot-per-class out- lier exposure OOD detection, although we do not include any OOD samples for our detection. Codes are available at https://github.com/lijingyao20010602/MOOD.

1. Introduction A reliable visual recognition system not only pro- vides correct predictions on known context (also known as in-distribution data) but also detects unknown out-of- distribution (OOD) samples and rejects (or transfers) them to human intervention for safe handling. This motivates ap- plications of outlier detectors before feeding input to the downstream networks, which is the main task of OOD de- tection, also referred to as novelty or anomaly detection. OOD detection is the task of identifying whether a test sam- ple is drawn far from the in-distribution (ID) data or not. It is at the cornerstone of various safety-critical applications, including medical diagnosis [5], fraud detection [45], au- tonomous driving [14], etc. CSI* ours889296100 89.294.9(a) One-class SSD+* ours949698100 94.697.6(b) Multi-class R50+ViT* ours96979899100 96.298.3(c) Near-Distribution R50+ViT* (1-shot)R50+ViT* (10-shot)ours (0-shot)99100 99.099.399.4(d) Few-shot Outlier ExposureFigure 1. Performance of MOOD compared with current SOTA (indicated by ‘*’) on four OOD detection tasks: (a) one-class OOD detection; (b) multi-class detection; (c) near-distribution detection; and (d) few-shot outlier exposure OOD detection. Many previous OOD detection approaches depend on outlier exposure [15, 53] to improve the performance of OOD detection, which turns OOD detection into a simple binary classification problem. We claim that the core of OOD detection is, instead, to learn the effective ID repre- sentation to discover OOD samples without any known out- lier exposure. In this paper, we first present our surprising finding – that is,simply using reconstruction-based methods can notably boost the performance on various OOD detection tasks . Our pioneer work along this line even outperforms previ- ous few-shot outlier exposure OOD detection, albeit we do not include any OOD samples. Existing methods perform contrastive learning [53,58] or pretrain classification on a large dataset [15] to detect OOD samples. The former methods classify images according to the pseudo labels while the latter classifies images based on ground truth, whose core tasks are both to fulfill the classifi- cation target. However, research on backdoor attack [50,51] shows that when learning is represented by classifying data, networks tend to take a shortcut to classify images. 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. 11578 In a typical backdoor attack scene [51], the attacker adds secret triggers on original training images with the visibly correct label. During the course of testing, the victim model classifies images with secret triggers into the wrong cate- gory. Research in this area demonstrates that networks only learn specific distinguishable patterns of different categories because it is a shortcut to fulfill the classification require- ment. Nonetheless, learning these patterns is ineffective for OOD detection since the network does not understand the intrinsic data distribution of the ID images. Thus, learning representations by classifying ID data for OOD detection may not be satisfying. For example, when the patterns sim- ilar to some ID categories appear in OOD samples, the net- work could easily interpret these OOD samples as the ID data and classify them into the wrong ID categories. To remedy this issue, we introduce the reconstruction- based pretext task. Different from contrastive learning in existing OOD detection approaches [53, 58], our method forces the network to achieve the training purpose of recon- structing the image and thus makes it learn pixel-level data distribution. Specifically, we adopt the masked image modeling (MIM) [2, 11, 20] as our self-supervised pretext task, which has been demonstrated to have great potential in both natu- ral language processing [11] and computer vision [2,20]. In the MIM task, we split images into patches and randomly mask a proportion of image patches before feeding the cor- rupted input to the vision transformer. Then we use the tokens from discrete V AE [47] as labels to supervise the network during training. With its procedure, the network learns information from remaining patches to speculate the masked patches and restore tokens of the original image. The reconstruction process enables the model to learn from the prior based on the intrinsic data distribution of images rather than just learning different patterns among categories in the classification process. In our extensive experiments, it is noteworthy that masked image modeling for OOD detection (MOOD) out- performs the current SOTA on all four tasks of one- class OOD detection, multi-class OOD detection, near- distribution OOD detection, and even few-shot outlier ex- posure OOD detection, as shown in Fig. 1. A few statistics are the following. 1. For one-class OOD detection (Tab. 6), MOOD boosts the AUROC of current SOTA, i.e., CSI [58], by 5.7% to94.9% . 2. For multi-class OOD detection (Tab. 7), MOOD out- performs current SOTA of SSD+ [53] by 3.0% and reaches 97.6% . 3. For near-distribution OOD detection (Tab. 2), AUROC of MOOD achieves 98.3% , which is 2.1% higher than the current SOTA of R50+ViT [15].4. For few-shot outlier exposure OOD detection (Tab. 9), MOOD ( 99.41% ) surprisingly defeats current SOTA of R50+ViT [15] (with 99.29% ), which makes use of 10 OOD samples per class. It is notable that we do not even include any OOD samples in MOOD.

Qin_Robust_3D_Shape_Classification_via_Non-Local_Graph_Attention_Network_CVPR_2023

Abstract We introduce a non-local graph attention network (NL- GAT), which generates a novel global descriptor through two sub-networks for robust 3D shape classification. In the first sub-network, we capture the global relationships between points ( i.e., point-point features) by designing a global relationship network (GRN). In the second sub- network, we enhance the local features with a geometric shape attention map obtained from a global structure net- work (GSN). To keep rotation invariant and extract more information from sparse point clouds, all sub-networks use the Gram matrices with different dimensions as input for working with robust classification. Additionally, GRN ef- fectively preserves the low-frequency features and improves the classification results. Experimental results on various datasets exhibit that the classification effect of the NLGAT model is better than other state-of-the-art models. Espe- cially, in the case of sparse point clouds ( 64points) with noise under arbitrary SO(3)rotation, the classification re- sult ( 85:4%) of NLGAT is improved by 39:4%compared with the best development of other methods.

1. Introduction 3D shape classification is one of the most critical tasks in 3D computer vision and computer graphics [7,10,18,37]. As 3D point cloud models are more accessible due to the rapid development of 3D scanning technology, their clas- sifications have attracted considerable attention in the last two decades [9, 14, 39]. The essential task for shape classification is to find a global descriptor for the input point cloud. Mainstream neu- ral networks have achieved excellent performance in point cloud classification on manually processed and aligned Corresponding author.data [15, 20, 25, 27, 35]. However, their performance tends to drop dramatically for complex real-world point clouds, which can be rotated (arbitrary orientation), sparse (with many missing parts), and noisy. Although there are methods for one or several states of complex point clouds classifica- tion through hand-crafted features, their global descriptors depend on the designed features [12, 29, 30, 38]. The reasons why current methods do not work well for complex point clouds are two folds. First, these methods tend to adopt aggregation operations of local features, by stacking hundreds of network layers as those in images [24], to obtain the global feature. Actually, it is difficult due to the point cloud network models using the point coordinates as input, and it will lead to feature homogenization, especially for the complex point clouds. Second, most of the meth- ods are not end-to-end and partially rely on the designed hand-crafted features, which can hardly capture the global information of the complex point clouds [2,5,31,32,36,41]. To this end, we propose an end-to-end deep learning network model built on complex point clouds, which con- sist of two global feature learning sub-networks for robust classification. In our model, we construct Gram matrices with different dimensions based on the input point coordi- nates for keeping rotation invariant, capturing crucial fea- tures (including local and non-local information with simi- lar structures) from noisy and sparse point clouds. The first sub-network based on multi-scale local Gram matrices is to extract the global relationships of point-point features in a shallow network layer through the network channel fusion operation ( i.e., channel attention mechanism). The second sub-network generates an attention map for enhancing the global relationships, from the global structure of a Gram matrix constructed by a whole point cloud. Finally, three fully connected (FC) layers receive the results learned on two sub-networks to generate a global descriptor for robust classification tasks. Contributions. Our contributions are summarized as fol- 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. 5374 lows. The global descriptor obtained by our method can well capture both the global relationship and global struc- ture, which outperforms existing methods in the task of classification for complex point clouds. We design an end-to-end deep learning network model, consisting of specific function modules in two global feature learning sub-networks. Our proposed modules, based on multi-scale Gram matrices constructed by the point coordinates, can gather lots of information for sparse point clouds, preserve valuable low-frequency features for noisy point clouds, and guarantee invari- ance to any rotational transformations.

Nair_Unite_and_Conquer_Plug__Play_Multi-Modal_Synthesis_Using_Diffusion_CVPR_2023

Abstract Generating photos satisfying multiple constraints finds broad utility in the content creation industry. A key hur- dle to accomplishing this task is the need for paired data consisting of all modalities (i.e., constraints) and their cor- responding output. Moreover, existing methods need re- training using paired data across all modalities to intro- duce a new condition. This paper proposes a solution to thisproblem based on denoising diffusion probabilistic models (DDPMs). Our motivation for choosing diffusion models over other generative models comes from the flexible in- ternal structure of diffusion models. Since each sampling step in the DDPM follows a Gaussian distribution, we show that there exists a closed-form solution for generating an image given various constraints. Our method can unite multiple diffusion models trained on multiple sub-tasks and conquer the combined task through our proposed sampling 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. 6070 strategy. We also introduce a novel reliability parame- ter that allows using different off-the-shelf diffusion mod- els trained across various datasets during sampling time alone to guide it to the desired outcome satisfying multi- ple constraints. We perform experiments on various stan- dard multimodal tasks to demonstrate the effectiveness of our approach. More details can be found at: https://nithin- gk.github.io/projectpages/Multidiff

1. Introduction Today’s entertainment industry is rapidly investing in content creation tasks [12, 22]. Studios and companies working on games or animated movies find various applica- tions of photos/videos satisfying multiple characteristics (or constraints) simultaneously. However, creating such photos is time-consuming and requires a lot of manual labor. This era of content creation has led to some exciting and valuable works like Stable Diffusion [28], Dall.E-2 [26], Imagen [29] and multiple other works that can create photorealistic im- ages using text prompts. All of these methods belong to the broad field of conditional image generation [25, 33]. This process is equivalent to sampling a point from the multi- dimensional space P(z|x)and can be mathematically ex- pressed as: ˆz∼P(z|x), (1) where ˆzdenotes the image to be generated based on a condi- tionx. The task of image synthesis becomes more restricted when the number of conditions increases, but it also hap- pens according to the user’s expectations. Several previous works have attempted to solve the conditional generation problem using generative models, such as V AEs [18,25] and Generative Adversarial Networks (GANs) [7, 34]. How- ever, most of these methods use only one constraint. In terms of image generation quality, the GAN-based meth- ods outperform V AE-based counterparts. Furthermore, dif- ferent strategies for conditioning GANs have been pro- posed in the literature. Among them, the text conditional GANs [3,27,41,45] embed conditional feature into the fea- tures from the initial layer through adaptive normalization scheme. For the case of image-level conditions such as a sketches or semantic labels, the conditional image is also the input to the discriminator and is embedded with an adap- tive normalization scheme [22, 31, 40, 42]. Hence, a GAN- based method for multimodal generation has multiple archi- tectural constraints [11] A major challenge in training generative models for mul- timodal image synthesis is the need for paired data con- taining multiple modalities [12, 32, 44]. This is one of the main reasons why most existing models restrict themselves to one or two modalities [32, 44]. Few works use more than two domain variant modalities for multimodal gener- ation [11, 45]. These methods can perform high-resolution Figure 2. An illustration of the difference between the existing multimodal generation approaches [45] and the proposed ap- proach. Existing multimodal methods require training on paired data across all modalities. In contrast, we present two ways that can be used for training: (1) Train with data pairs belonging to dif- ferent modalities one at a time, and (2) Train only for the additional modalities using a separate diffusion model in case existing mod- els are available for the remaining modalities. During sampling, we forward pass for each conditioning strategy independently and combine their corresponding outputs, hence preserving the differ- ent conditions. image synthesis and require training with paired data across different domains to achieve good results. But to increase the number of modalities, the models need to be retrained; thus they do not scale easily. Recently, Shi et al. [32] pro- posed a weakly supervised V AE-based multimodal genera- tion method without paired data from all modalities. The model performs well when trained with sparse data. How- ever, if we need to increase the number of modalities, the model needs to be retrained; therefore, it is not scalable. Scalable multimodal generation is an area that has not been properly explored because of the difficulty in obtaining the large amounts of data needed to train models for the gener- ative process. Recently diffusion models have outperformed other gen- erative models in the task of image generation [5, 9]. This is due to the ability of diffusion models to perform exact sampling from very complex distributions [33]. A unique quality of the diffusion models compared to other genera- tive processes is that the model performs generation through a tractable Markovian process, which happens over many time steps. The output at each timestep is easily accessible. Therefore, the model is more flexible than other generative models, and this form of generation allows manipulation of images by adjusting latents [1, 5, 23]. Various techniques have used this interesting property of diffusion models for low-level vision tasks such as image editing [1, 17], im- age inpainting [20], image super-resolution [4], and image restoration problems [16]. In this paper, we exploit this flexible property of the de- noising diffusion probabilistic models and use it to design a solution to multimodal image generation problems with- 6071 Figure 3. An illustration of our proposed approach. During training, we use diffusion models trained across multiple datasets (we can either train a single model that supports multiple different conditional strategies one at a time or multiple models). During Inference, we sample using the proposed approach and condition them using different modalities at the same time. out explicitly retraining the network with paired data across all modalities. Figure 2 depicts the comparison between existing methods and our proposed method. Current ap- proaches face a major challenge: the inability to combine models trained across different datasets during inference time [9, 19]. In contrast, our work allows users flexibility during training and can also use off-the-shelf models for multi-conditioning, providing greater flexibility when us- ing diffusion models for multimodal synthesis task. Figure 1 visualizes some applications of our proposed approach. As shown in Figure 1-(a), we use two open-source mod- els [5,21] for generic scene creation. Using these two mod- els, we can bring new novel categories into an image (e.g. Otterhound: the rarest breed of dog). We also illustrate the results showing multimodal face generation, where we use a model trained to utilize different modalities from differ- ent datasets. As it can be seen in 1-(b) and (c), our work can leverage models trained across different datasets and combine them for multi-conditional synthesis during sam- pling. We evaluate the performance of our method for the task of multimodal synthesis using the only existing mul- timodal dataset [45] for face generation where we condi- tion based on semantic labels and text attributes. We also evaluate our method based on the quality of generic scene generation. The main contributions of this paper are summarized as follows: • We propose a diffusion-based solution for image gen-eration under the presence of multimodal priors. • We tackle the problem of need for paired data for mul- timodal synthesis by deriving upon the flexible prop- erty of diffusion models. • Unlike existing methods, our method is easily scalable and can be incorporated with off-the-shelf models to add additional constraints.

Li_Neuralangelo_High-Fidelity_Neural_Surface_Reconstruction_CVPR_2023

Abstract Neural surface reconstruction has been shown to be pow- erful for recovering dense 3D surfaces via image-based neu- ral rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo , which combines the rep- resentation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our ap- proach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neu- ralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpass- ing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures.

1. Introduction 3D surface reconstruction aims to recover dense geomet- ric scene structures from multiple images observed at differ- ent viewpoints [9]. The recovered surfaces provide structural information useful for many downstream applications, such as 3D asset generation for augmented/virtual/mixed real- ity or environment mapping for autonomous navigation of robotics. Photogrammetric surface reconstruction using a monocular RGB camera is of particular interest, as it equips users with the capability of casually creating digital twins of the real world using ubiquitous mobile devices. Classically, multi-view stereo algorithms [6, 16, 29, 34] had been the method of choice for sparse 3D reconstruc- tion. An inherent drawback of these algorithms, however, is their inability to handle ambiguous observations, e.g. regions with large areas of homogeneous colors, repetitive texture 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. 8456 patterns, or strong color variations. This would result in inaccurate reconstructions with noisy or missing surfaces. Recently, neural surface reconstruction methods [36, 41, 42] have shown great potential in addressing these limitations. This new class of methods uses coordinate-based multi-layer perceptrons (MLPs) to represent the scene as an implicit function, such as occupancy fields [25] or signed distance functions (SDF) [36, 41, 42]. Leveraging the inherent con- tinuity of MLPs and neural volume rendering [22], these techniques allow the optimized surfaces to meaningfully in- terpolate between spatial locations, resulting in smooth and complete surface representations. Despite the superiority of neural surface reconstruction methods over classical approaches, the recovered fidelity of current methods does not scale well with the capacity of MLPs. Recently, Müller et al. [23] proposed a new scalable representation, referred to as Instant NGP (Neural Graphics Primitives). Instant NGP introduces a hybrid 3D grid struc- ture with a multi-resolution hash encoding and a lightweight MLP that is more expressive with a memory footprint log- linear to the resolution. The proposed hybrid representation greatly increases the representation power of neural fields and has achieved great success at representing very fine- grained details for a wide variety of tasks, such as object shape representation and novel view synthesis problems. In this paper, we propose Neuralangelo for high-fidelity surface reconstruction (Fig. 1). Neuralangelo adopts In- stant NGP as a neural SDF representation of the underlying 3D scene, optimized from multi-view image observations via neural surface rendering [36]. We present two findings central to fully unlocking the potentials of multi-resolution hash encodings. First, using numerical gradients to compute higher-order derivatives, such as surface normals for the eikonal regularization [8, 12, 20, 42], is critical to stabilizing the optimization. Second, a progressive optimization sched- ule plays an important role in recovering the structures at different levels of details. We combine these two key ingredi- ents and, via extensive experiments on standard benchmarks and real-world scenes, demonstrate significant improvements over image-based neural surface reconstruction methods in both reconstruction accuracy and view synthesis quality. In summary, we present the following contributions: •We present the Neuralangelo framework to naturally incorporate the representation power of multi-resolution hash encoding [23] into neural SDF representations. •We present two simple techniques to improve the quality of hash-encoded surface reconstruction: higher-order derivatives with numerical gradients and coarse-to-fine optimization with a progressive level of details. •We empirically demonstrate the effectiveness of Neu- ralangelo on various datasets, showing significant im- provements over previous methods.

Luo_Class-Incremental_Exemplar_Compression_for_Class-Incremental_Learning_CVPR_2023

Abstract Exemplar-based class-incremental learning (CIL) [36] finetunes the model with all samples of new classes but few- shot exemplars of old classes in each incremental phase, where the “few-shot” abides by the limited memory bud- get. In this paper, we break this “few-shot” limit based on a simple yet surprisingly effective idea: compressing exemplars by downsampling non-discriminative pixels and saving “many-shot” compressed exemplars in the mem- ory. Without needing any manual annotation, we achieve this compression by generating 0-1masks on discrimina- tive pixels from class activation maps (CAM) [49]. We propose an adaptive mask generation model called class- incremental masking (CIM) to explicitly resolve two dif- ficulties of using CAM: 1) transforming the heatmaps of CAM to 0-1masks with an arbitrary threshold leads to a trade-off between the coverage on discriminative pix- els and the quantity of exemplars, as the total memory is fixed; and 2) optimal thresholds vary for different object classes, which is particularly obvious in the dynamic envi- ronment of CIL. We optimize the CIM model alternatively with the conventional CIL model through a bilevel opti- mization problem [40]. We conduct extensive experiments on high-resolution CIL benchmarks including Food-101, ImageNet-100, and ImageNet-1000, and show that using the compressed exemplars by CIM can achieve a new state- of-the-art CIL accuracy, e.g., 4.8 percentage points higher than FOSTER [42] on 10-Phase ImageNet-1000. Our code is available at https://github.com/xfflzl/CIM-CIL.

1. Introduction Dynamic AI systems have a continual learning nature to learn new class data. They are expected to adapt to new classes while maintaining the knowledge of old classes, i.e., free from forgetting problems [31]. To evaluate this, the following protocol of class-incremental learning (CIL) was proposed by Rebuffi et al. [36]. The model training goes Phase i herding herdingPhase i+1 Phase i herding distillingPhase i+1 all samples Phase i+1 Phase i herding New Data Old Exemplars New Data Old Exemplars Original Images in (a) JPEG Images in (c) Distilled Images in (b) JPEG herding samplesPhase i herding maskingPhase i+1 herding samplesNew Data Old Exemplars New Data Old Exemplars Masked Images in (d) (a) iCaRL [baseline] (b) Mnemonics [related] (c) MRDC [related] (d) CIM-based CIL [ours] New Data Old Exemplars New Data Old Exemplars New Data Old Exemplars New Data Old Exemplars Figure 1. The phase-wise training data in different methods. (a) iCaRL [36] is the baseline method using full new class data and few-shot old class exemplars. (b) Mnemonics [27] distills all training samples into few-shot exemplars without increasing their quantity. (c) MRDC [43] compresses each exemplar uniformly into a low-resolution image using JPEG [41]. (d) Our approach based on the proposed class-incremental masking (CIM) down- samples only non-discriminative pixels in the image. The legend shows the symbols of special images generated by the methods. through a number of phases. Each phase has new class data added and old class data discarded, and the resultant model is evaluated on the test data of all seen classes. A straightforward way to retain old class knowledge is keep- ing around a few old class exemplars in the memory and using them to re-train the model in subsequent phases. The number of exemplars is usually limited, e.g., 5∼20exem- plars per class [12, 17, 25, 27, 36, 42, 44, 46, 48], as the total memory in CIL strictly budgeted, e.g., 2kexemplars. This leads to a serious data imbalance between old and new classes, e.g., 20per old class vs.1.3kper new class (on ImageNet-1000 [9]), as illustrated in Figure 1a. The train- ing is thus always dominated by new classes, and forgetting problems occur for old classes. Liu et al. [27] tried to miti- gate this problem by parameterizing and distilling the exem- 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. 11371 plars, without increasing the number of them (Figure 1b). Wang et al. [43] traded off between the quality and quantity of exemplars by uniformly compressing exemplar images with JPEG [41] (Figure 1c). As shown in Figure 1d, our approach is also based on image compression. The idea is to downsample only non-discriminative pixels (e.g., back- ground) and keep discriminative pixels (i.e., representative cues of foreground objects) as the original. In this way, we do not sacrifice the discriminativeness of exemplars when increasing their quantity. In particular, we aim for adaptive compression in dynamic environments of CIL, where the intuition is later phases need to be more conservative (i.e., less downsampling) as the model needs more visual cues to classify the increased number of classes. To achieve selective and adaptive compression, we need the location labels of discriminative pixels. Without extra labeling, we automatically generate the labels by utilizing the model’s own “attention” on discriminative features, i.e., class activation maps (CAM) [49]. We take this method as a feasible baseline, and based on it, we propose an adaptive version called class-incremental masking (CIM). Specifi- cally, for each input image (with its class label), we use its feature maps and classifier weights (corresponding to its class label) to compute a CAM by channel-wise mul- tiplication, aggregation, and normalization. Then, we ap- ply hard thresholding to generate a 0-1mask.1We notice that when generating the masks in the dynamic environ- ments of CIL, the optimal hyperparameters (such as the value of hard threshold and the choice of activation func- tions) vary for different classes as well as in different incre- mental phases. Our adaptive version CIM tackles this by pa- rameterizing a mask generation model and optimizing it in an end-to-end manner across all incremental phases. In each phase, the learned CIM model adaptively generates class- and phase-specific masks. We find that the compressed ex- emplars based on these masks have stronger representative- ness, compared to using the conventional CAM. Technically, we have two models to optimize, i.e., the CIL model and the CIM model.2These two cannot be optimized separately as they are dependent on computa- tion: 1) the CIM model compresses exemplars to input into the CIL model; 2) the two models share network pa- rameters. We exploit a global bilevel optimization prob- lem (BOP) [7, 40] to alternate their training processes at two levels. This BOP goes through all incremental train- ing phases. In particular, for each phase, we perform a lo- cal BOP with two steps to tune the parameters of the CIM model: 1) a temporary model is trained with the compressed exemplars as input; and 2) a validation loss on the uncom- 1Note that we do not use mask labels to do image compression because storing them is expensive. Instead, we expand the mask to a bounding box, as elaborated in Section 4. 2Note that the CIM model is actually a plug-in branch in the CIL model, which is detailed in Section 4.2.pressed new data is computed and the gradients are back- propagated to optimize the parameters of CIM. To evalu- ate CIM, we conduct extensive experiments by plugging it in recent CIL methods,3LUCIR [17], DER [46], and FOSTER [42], on three high-resolution benchmarks, Food- 101 [3], ImageNet-100 [17], and ImageNet-1000 [9]. We find that using the compressed exemplars by CIM brings consistent and significant improvements, e.g., 4.2%and 4.8%higher than the SOTA method FOSTER [42], respec- tively, in the 5-phase and 10-phase settings of ImageNet- 1000, with a total memory budget for 5kexemplars.

Oh_BlackVIP_Black-Box_Visual_Prompting_for_Robust_Transfer_Learning_CVPR_2023

Abstract With the surge of large-scale pre-trained models (PTMs), fine-tuning these models to numerous downstream tasks be- comes a crucial problem. Consequently, parameter effi- cient transfer learning (PETL) of large models has grasped huge attention. While recent PETL methods showcase im- pressive performance, they rely on optimistic assumptions: 1) the entire parameter set of a PTM is available, and 2) a sufficiently large memory capacity for the fine-tuning is equipped. However, in most real-world applications, PTMs are served as a black-box API or proprietary software with- out explicit parameter accessibility. Besides, it is hard to meet a large memory requirement for modern PTMs. In this work, we propose black-box visual prompting (Black- VIP), which efficiently adapts the PTMs without knowl- edge about model architectures and parameters. Black- VIP has two components; 1) Coordinator and 2) simul- taneous perturbation stochastic approximation with gradi- ent correction (SPSA-GC). The Coordinator designs input- dependent image-shaped visual prompts, which improves few-shot adaptation and robustness on distribution/location shift. SPSA-GC efficiently estimates the gradient of a tar- get model to update Coordinator. Extensive experiments on 16 datasets demonstrate that BlackVIP enables robust adaptation to diverse domains without accessing PTMs’ parameters, with minimal memory requirements. Code: https://github.com/changdaeoh/BlackVIP

1. Introduction Based on their excellent transferability, large-scale pre- trained models (PTMs) [7, 17, 54] have shown remarkable success on tasks from diverse domains and absorbed in- creasing attention in machine learning communities. By witnessing PTMs’ success, Parameter-Efficient Transfer Learning (PETL) methods that efficiently utilize the PTMs †Work done at University of Seoul ‡Corresponding author; Work partly done at University of Seoul Figure 1. While FT updates the entire model, VP has a small num- ber of parameters in the input pixel space. However, VP still re- quires a large memory capacity to optimize the parameters through backpropagation. Moreover, FT and VP are only feasible if the PTM’s parameters are accessible. Meanwhile, BlackVIP does not assume the parameter-accessibility by adopting a black-box opti- mization (SPSA-GC) algorithm rather than relying on backpropa- gation. Besides, BlackVIP reparameterizes the visual prompt with a neural network and optimizes tiny parameters with SPSA-GC. Based on the above properties, BlackVIP can be widely adopted in realistic and resource-limited transfer learning scenarios. are recently emerging. While the standard fine-tuning (FT) and its advanced variants [38, 73] update the entire or large portion of a PTM [16], PETL methods aim to achieve com- parable performance to FT by optimizing a small number of learnable parameters. Among them, prompt-based approaches [3,5,33,40,41] have been widely investigated from diverse research areas. For vision PTMs, Visual Prompt Tuning [33] injects a few additional learnable prompt tokens inside of ViT’s [17] lay- ers or embedding layer and only optimizes them. Bahng et al. [3] investigate visual prompting (VP), which adopts the learnable parameters on input pixel space as a visual prompt, while no additional modules are inserted into the pre-trained visual model. Besides, prompt learning meth- ods for VLM are also actively studied [35, 78, 81, 83]. While existing PETL methods show impressive perfor- mance with few learnable parameters, they rely on two 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. 24224 optimistic assumptions. First, the previous PETL as- sumes that the full parameters of the PTM are accessible. However, many real-world AI applications are served as API and proprietary software, and they do not reveal the implementation-level information or full parameters due to commercial issues, e.g., violating model ownership. As a result, exploiting high-performing PTMs to specific down- stream tasks not only in the white-box setting but also black-box setting (limited accessibility to the model’s de- tail) is a crucial but unexplored problem. Second, exist- ing methods require a large memory capacity. While PETL approaches have few learnable parameters, they require a large amount of memory for backpropagating the gradient throughout the large-scale PTM parameters to learnable pa- rameters. Therefore, users who want to adopt a large-scale PTM should satisfy large memory requirements despite the small learnable parameters. Besides, if the users entrust PTM fine-tuning to the model owner with their specific data, data-privacy concerns will inevitably arise [74]. To alleviate the above unrealistic assumptions, we are pioneering black-box visual prompting (BlackVIP) approach, which enables the parameter-efficient transfer learning of pre-trained black-box vision models from the low-resource user perspective (illustrated in Figure 1). BlackVIP works based on the following two core compo- nents: 1) pixel space input-dependent visual prompting and 2) a stable zeroth-order optimization algorithm. Firstly, we augment an input image by attaching an vi- sual prompt per pixel. It is noted that input space prompt- ing does not require the accessibility on parts of architec- ture [37, 78] or the first embedding layer [35, 81, 83] of PTM. While the previous works only introduce a pixel-level prompt to a small fraction of the fixed area, such as out- side of the image [3], BlackVIP designs the prompt with the same shape as the original given image to cover the entire image view. Therefore, our prompt has a higher capability and can flexibly change the semantics of the original image. In addition, we reparameterize the prompt with a neural net- work. Specifically, we propose the Coordinator , an asym- metric autoencoder-style network that receives the original image and produces a corresponding visual prompt for each individual image. As a result, Coordinator automatically designs each prompt conditioned on the input rather than the shared manual design of a previous work [3]. By opti- mizing the reparameterized model instead of the prompt it- self, we greatly reduce the number of parameters (from 69K of VP [3] to 9K) so that suitable for black-box optimization. Next, unlike other PETL approaches, BlackVIP adopts a zeroth-order optimization (ZOO) that estimates the zeroth- order gradient for the coordinator update to relax the as- sumption that requires access to the huge PTM parame- ters to optimize the prompt via backpropagation. There- fore, BlackVIP significantly reduces the required mem-ory for fine-tuning. Besides, we present a new ZOO al- gorithm, Simultaneous Perturbation Stochastic Approx- imation with Gradient Correction (SPSA-GC) based on (SPSA) [58]. SPSA-GC first estimates the gradient of the target black-box model based on the output difference of perturbed parameters and then corrects the initial estimates in a momentum-based look-ahead manner. By integrating the Coordinator and SPSA-GC, BlackVIP achieves signifi- cant performance improvement over baselines. Our main contributions are summarized as follows: • To our best knowledge, this is the first paper that ex- plores the input-dependent visual prompting on black- box settings. For this, we devise Coordinator, which reparameterizes the prompt as an autoencoder to han- dle the input-dependent prompt with tiny parameters. • We propose a new ZOO algorithm, SPSA-GC, that gives look-ahead corrections to the SPSA’s estimated gradient resulting in boosted performance. • Based on Coordinator and SPSA-GC, BlackVIP adapts the PTM to downstream tasks without parame- ter access and large memory capacity. We extensively validate BlackVIP on 16 datasets and demonstrate its effectiveness regarding few-shot adaptability and ro- bustness on distribution/object-location shift.

Noh_Disentangled_Representation_Learning_for_Unsupervised_Neural_Quantization_CVPR_2023

Abstract The inverted index is a widely used data structure to avoid the infeasible exhaustive search. It accelerates re-trieval significantly by splitting the database into multipledisjoint sets and restricts distance computation to a smallfraction of the database. Moreover , it even improves searchquality by allowing quantizers to exploit the compact distri-bution of residual vector space. However , we firstly pointout a problem that an existing deep learning-based quan- tizer hardly benefits from the residual vector space, unlike conventional shallow quantizers. To cope with this problem,we introduce a novel disentangled representation learningfor unsupervised neural quantization. Similar to the con-cept of residual vector space, the proposed method enablesmore compact latent space by disentangling information of the inverted index from the vectors. Experimental results onlarge-scale datasets confirm that our method outperforms the state-of-the-art retrieval systems by a large margin.

1. Introduction Measuring the distances among feature vectors is a fun- damental requirement in various fields of computer vision.One of the tasks most relevant to distance measurement is the nearest neighbor search, which finds the closest data in the database from a query. The task is especially chal- lenging in high-dimensional and large-scale databases dueto huge computational costs and memory overhead. By relaxing the complexity, Approximate Nearest Neighbor (ANN) search is popular in practice. Recent ap-proaches for ANN typically learn the compact representa- tion by exploiting Multi-Codebook Quantization (MCQ) [ 2, 9,16]. Compared to hashing-based approaches [ 1,10,12], the MCQ provides a more informative asymmetric distanceestimator where the query side is not compressed. More-over, all possible distances between the query and code-words can be stored in a lookup table for efficiency. Although the MCQ accelerates the distance computation with the lookup table, exhaustive search on the large-scale *Corresponding authordataset is still prohibited. The Inverted File with Asym-metric Distance Computation (IVFADC) [ 16] is proposed for non-exhaustive ANN search by cooperating with the in-verted index [ 30]. It splits the database into multiple disjoint sets and restricts distance computations to small portions close to the query to accelerate the retrieval speed. More-over, the compactness of residual vector space between data points and inverted indices substantially enhances the quan-tization quality. Thanks to the rapid advances in deep learning, most ar- eas of computer vision benefit from its great learning capac- ity compared to shallow methods. However, the state-of- the-art methods of unsupervised quantization remain shal- low for a long time because selecting the maximum value (i.e. argmax), which is an essential operation of quanti-zation, is not differentiable. Inspired by a recent gener- ative model with discrete hidden variables [ 35], the Un- supervised Neural Quantization (UNQ) [ 23] introduced an encoder-decoder-based architecture for ANN search. Thelarge learning capacity of deep neural architecture signifi- cantly improves the retrieval quality compared to conven- tional shallow methods. Despite the outperforming performance of the UNQ, its superiority is validated only on the exhaustive search. Toverify its effectiveness on non-exhaustive search, we con- duct an experiment of non-exhaustive UNQ with an inverted index. Interestingly, we observe that this deep architecturedoes not benefit from the residual vector space and it evenharms the search quality as reported in Table 1. We hypoth- esize the reasons for this performance degradation from twoperspectives. First, both the residual vector space and la-tent space of the neural network transform the data into aquantization-friendly distribution, thus deep quantizer hasa scant margin to be improved by the residual space. Sec-ond, residual space sacrifices the distributional characteris-tics of each cluster, since the information of cluster centerin the original space is removed. For conventional shallowquantizers, the drawback of residual space is obscured byits huge advantage of making a compact distribution. How-ever, deep quantizer only takes the disadvantages (informa-tion loss) from residual space without leveraging the effec- 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. 12001 tiveness such as compactness of residuals. In this paper, we focus on extending the application of deep architectures for non-exhaustive search. To thisend, we learn a disentangled representation to harmonizea deep architecture with the inverted index, inspired by re-cent representation learning techniques for generative mod- els [ 8,34]. In our disentangled representation learning, both encoder and decoder get information of cluster center as anadditional input. Since the information of cluster center is redundant to decoder if latent feature contains information of cluster center, the encoder is trained to remove the infor- mation of cluster centers from the latent embedding. Thedisentangled representation learning is similar to concept ofthe residual vector space that provides more compact dis- tribution by taking out the information of cluster centers.The experimental results verify that the learning disentan- gled representation enables the neural quantization to col- laborate with inverted index and outperforms the state-of-the-art methods. The contributions of our paper include: • We point out that the residual encoding of the inverted index is incompatible with the neural multi-codebookquantization method. • We propose a novel disentangled representation learn- ing for neural multi-codebook quantization to combinedeep quantization and inverted index. • The experimental results show that the proposed method outperforms the state-of-the-art retrieval sys- tems by a large margin.

Olber_Detection_of_Out-of-Distribution_Samples_Using_Binary_Neuron_Activation_Patterns_CVPR_2023

Abstract Deep neural networks (DNN) have outstanding perfor- mance in various applications. Despite numerous efforts of the research community, out-of-distribution (OOD) samples remain a significant limitation of DNN classifiers. The abil- ity to identify previously unseen inputs as novel is crucial in safety-critical applications such as self-driving cars, un- manned aerial vehicles, and robots. Existing approaches to detect OOD samples treat a DNN as a black box and evalu- ate the confidence score of the output predictions. Unfortu- nately, this method frequently fails, because DNNs are not trained to reduce their confidence for OOD inputs. In this work, we introduce a novel method for OOD detection. Our method is motivated by theoretical analysis of neuron ac- tivation patterns (NAP) in ReLU-based architectures. The proposed method does not introduce a high computational overhead due to the binary representation of the activation patterns extracted from convolutional layers. The extensive empirical evaluation proves its high performance on vari- ous DNN architectures and seven image datasets.

1. Introduction Even the most efficient deep neural network (DNN) ar- chitectures, designed for image recognition tasks, cannot ensure that they will not malfunction during their opera- tion. Thus, deployment of those safety-critical applications, such as in self-driving cars, unmanned aerial vehicles, and robots, is still an unresolved problem [9, 35]. The use of safety mechanisms, such as runtime monitors, is a viable strategy to keep the system in a safe state despite of DNN failure. The design and development of such monitors in the context of safety-critical applications is a significant chal- lenge [10,48]. Therefore, it is required to define robust met- rics that can allow to detect and control of DNN’s failures at runtime and mitigate potential hazards caused by their performance limitations. DNNs are trained over a set of inputs sampled from real-world scenarios. However, due to the large variation of the input images, the training dataset cannot contain all possi- ble variants of input samples. Although it is expected that the trained models can perform well on unknown inputs, es- pecially those that are similar to the training data, it cannot be guaranteed that they will perform well for OOD noisy samples that present the objects not considered before [15]. While DNN training techniques should allow a network to achieve high generalization capabilities, it is also crucial to ensure the dependability of safety-critical systems to train a model so that any outlying input will result in low confi- dence of the network’s decision. The fundamental challenge to ensure the safety of DNNs is to estimate if a given input sample comes from the same data distribution for which a DNN was trained. This is very hard to estimate because the network usually extrapolates its decision while receiving new image samples. Another cause of the incorrect recognition of outlying samples can be the distributional shift of input data over time (e.g. the time-dependent variations of an object’s appearance) [19]. In the vast literature, this problem has been formulated as a problem of detecting whether input data are from an in-distribution (ID) or out-of-distribution (OOD). This has been studied for many years and discussed in the follow- ing aspects: sample rejection, anomaly detection, open-set samples recognition, familiar vs unfamiliar samples or un- certainty estimation [2, 24, 38]. In this work, we present a novel algorithm for the identi- fication of OOD image samples. In our method, we extract the binary neuron activation patterns on various hidden lay- ers of a DNN and compare them with the ones collected in the training procedure. By measuring the Hamming dis- tances between extracted binary patterns of any test sample and the patterns extracted during the training, we can iden- tify OOD samples. The main contributions of this paper are the following: • We introduce NAP - an algorithm that extracts bi- nary patterns from both fully connected and convolu- tional layers and estimates a classifier’s predictive un- 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. 3378 certainty based on the patterns. The proposed method outperforms state-of-the-art OOD detection methods. Moreover, the algorithm is straightforward, making it simple to incorporate into existing DNN architectures. • We provide an extended empirical evaluation compar- ing the impact of the activation patterns collected from different layers of DNN which may inspire future re- search in this area. • We publish the largest evaluation framework for OOD detection. This framework contains 17 OOD methods (including the proposed NAP-based method) that can be directly tested on two state-of-the-art DNN archi- tectures and 7 datasets allowing for simple extension of the framework for new methods, architectures, and datasets.1

Mu_Progressive_Backdoor_Erasing_via_Connecting_Backdoor_and_Adversarial_Attacks_CVPR_2023

Abstract Deep neural networks (DNNs) are known to be vulnera- ble to both backdoor attacks as well as adversarial attacks. In the literature, these two types of attacks are commonly treated as distinct problems and solved separately, since they belong to training-time and inference-time attacks re- spectively. However, in this paper we find an intriguing con- nection between them: for a model planted with backdoors, we observe that its adversarial examples have similar be- haviors as its triggered images, i.e., both activate the same subset of DNN neurons. It indicates that planting a back- door into a model will significantly affect the model’s ad- versarial examples. Based on these observations, a novel Progressive Backdoor Erasing (PBE) algorithm is proposed to progressively purify the infected model by leveraging un- targeted adversarial attacks. Different from previous back- door defense methods, one significant advantage of our ap- proach is that it can erase backdoor even when the clean extra dataset is unavailable. We empirically show that, against 5 state-of-the-art backdoor attacks, our PBE can effectively erase the backdoor without obvious performance degradation on clean samples and outperforms existing de- fense methods.

1. Introduction Deep neural networks (DNNs) have been widely adopted in many safety-critical applications ( e.g., face recognition and autonomous driving), thus more attention has been paid to the security of deep learning. It has been demonstrated that DNNs are prone to potential threats in both their infer- ence as well as training phases. Inference-time attack (a.k.a. adversarial attack [5, 25]) aims to fool a trained model into making incorrect predictions with small adversarial pertur- bations. In contrast, training-time attack (a.k.a. backdoor attack [13]) attempts to plant a backdoor into a model in the training phase, so that the infected model would mis- classify the testing images as the target-label whenever a pre-defined trigger (e.g., several pixels) is embedded into them ( i.e., triggered testing images). 0 1 2 3 4 5 6 7 8 9 Predicted label0 1 2 3 4 5 6 7 8 9 Target label0.17 0.33 0.03 0.05 0.01 0.02 0.02 0.03 0.18 0.16 0.04 0.06 0.37 0.06 0.10 0.01 0.03 0.04 0.23 0.07 0.07 0.26 0.03 0.03 0.01 0.02 0.02 0.01 0.13 0.43 0.19 0.00 0.08 0.16 0.20 0.10 0.21 0.05 0.01 0.01 0.04 0.01 0.14 0.01 0.14 0.41 0.16 0.06 0.02 0.03 0.06 0.00 0.24 0.21 0.03 0.21 0.10 0.14 0.01 0.01 0.01 0.01 0.18 0.44 0.12 0.03 0.08 0.11 0.01 0.01 0.02 0.03 0.27 0.31 0.16 0.06 0.07 0.01 0.03 0.06 0.08 0.01 0.09 0.10 0.23 0.30 0.02 0.13 0.01 0.04 0.45 0.15 0.06 0.04 0.03 0.02 0.04 0.01 0.07 0.13 0.00.10.20.30.40.50.6(a) For benign model. 0 1 2 3 4 5 6 7 8 9 Predicted label0 1 2 3 4 5 6 7 8 9 Target label0.38 0.05 0.10 0.10 0.07 0.10 0.04 0.05 0.05 0.07 0.06 0.43 0.09 0.09 0.06 0.07 0.05 0.04 0.04 0.06 0.06 0.04 0.49 0.09 0.06 0.08 0.05 0.05 0.03 0.05 0.06 0.04 0.08 0.52 0.06 0.08 0.04 0.04 0.03 0.05 0.06 0.04 0.10 0.09 0.46 0.08 0.05 0.04 0.03 0.05 0.06 0.04 0.09 0.09 0.07 0.47 0.04 0.05 0.03 0.05 0.05 0.04 0.10 0.09 0.06 0.07 0.45 0.04 0.03 0.05 0.06 0.04 0.09 0.09 0.06 0.08 0.04 0.46 0.03 0.05 0.07 0.04 0.08 0.09 0.06 0.07 0.04 0.04 0.44 0.06 0.06 0.05 0.08 0.09 0.06 0.08 0.04 0.04 0.04 0.46 0.10.20.30.40.5 (b) For infected model. Figure 1. Predicted labels v.s. Target-labels for 10,000randomly sampled adversarial examples from CIFAR-10, with respect to benign andinfected models. (a) For a benign model, the predicted labels obey uniform distribution; (b) for infected models under WaNet backdoor attack [20], its adversarial examples are highly likely to be classified as the target-label (the matrix diagonals). Figure 2. Illustration of our observations. For benign models, conducting an untargeted adversarial attack will make an image move close to anyclass ( e.g., Class 0or Class 2) in feature space. But for infected models, adversarial attack will make it move close to the target-label class ( e.g., Class l) Due to the obvious differences between backdoor and adversarial attacks, they are often treated as two different problems and solved separately in the literature. But in this paper, we illustrate that there is an underlying connection between them, i.e., planting a backdoor into one model will significantly affect the model’s adversarial examples. More- over, based on such findings we propose a new method to defend against backdoor attacks by leveraging adversarial attack techniques ( i.e., generating adversarial examples). 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. 20495 In particular, we observe that: for a model planted with backdoors, its adversarial examples have similar behaviors as its triggered images. This is significantly different from a benign model without backdoors. Specifically, for a benign model , the predicted class labels of its adversarial examples obey a uniform distribution, as shown in Fig.1a. However, for an infected model , we surprisingly observe that its ad- versarial examples are highly likely to be predicted as the backdoor target-label , as shown in Fig.1b. As we know, triggered images will also be predicted as the backdoor target-label by an infected model. Therefore, it means that adversarial examples have similar behaviors as its triggered images for an infected model. Particularly, these phenom- ena are present regardless of the target-label, the backdoor attack setting ( i.e., all-to-one or all-to-all settings), and even for most trigger embedding mechanisms ( e.g., adding [6], blending [3] or warping [20]). To find the underlying reason of such phenomena, we measure the feature similarity of those adversarial images and triggered images. Briefly, we find that after planting a backdoor into one model, the features of adversarial images change significantly. Particularly, the features of adversar- ial image ex′are surprisingly very similar to that of triggered image xt, as illustrated in Fig.2 and Fig.3. It indicates that both the ex′andxthave similar behaviors, i.e.,both ac- tivate the same subset of DNN neurons . Note that such connection between adversarial and backdoor attack could be leveraged to design backdoor defense methods. Backdoor attacks made great advances in recent years, evolved from visible trigger [6] to invisible trigger [3, 16, 20], from poisoning label to clean-label attacks [1]. For ex- ample, WaNet [20] uses affine transformation as trigger em- bedding mechanism, which could significant improve the invisibility of trigger. In contrast, the research on backdoor defenses lag behind a little. Even for the state-of-the-art backdoor defense methods [12,14,17], most of them can be evaded by the advanced modern backdoor attacks. More- over, a clean extra dataset is often required by those defense methods to erase backdoor from infected models. In this paper, we propose a new backdoor defense method based on the discovered connections between ad- versarial and backdoor attacks, which could not only de- fend against modern backdoor attacks but also work with- out a clean extra dataset. Specifically, at the beginning the training data (containing poisoning images) are randomly sampled to build an initial extra dataset. Next, we use them to purify the infected model by leveraging adversarial at- tack techniques. And then, the purified model is used to identify clean images from training data, which are used to update the extra dataset. With an alternating procedure, the infected model as well as the extra dataset are progressively purified. So, we call our approach Progressive Backdoor Erasing (PBE).Regarding how to purify the infected model, we gener- ate adversarial examples and use them to fine-tune the in- fected model. Since adversarial images could come from arbitrary class, such fine-tuning procedure works like asso- ciating triggered images to arbitrary class instead of just the target class, which breaks the foundation of backdoor at- tacks ( i.e., building a strong correlation between a trigger pattern and a target-label [12]). That is why our approach can erase backdoor from infected models. As for identifying clean images, since clean images have similar prediction results for both benign and infected mod- els, we could effectively identify them by using the previ- ously obtained purified model. Note that if a clean extra dataset is available, we can skip the step of purifying extra dataset, and only run the step of purifying model once. A big advantage of our approach is that it does not need the clean extra dataset and it can progressively filter poi- soning training data to obtain clean data. In our approach, the purified model could help to obtain clean data, in return the obtained clean data could help to further purify model. Thus, the alternating iterations could progressively improve each other. To the best of knowledge, our approach is the first work to defend against backdoor attack without a clean extra dataset. Our main contributions are summarized as follows: •We observe an underlying connection between back- door attacks and adversarial attacks, i.e., for an in- fected model, its adversarial examples have similar be- haviors as its triggered samples. And an theoretical analysis is given to justify our observation. •According to our observations, we propose a progres- sive backdoor defense method, which achieves the state-of-the-art defensive performance, even when a clean extra dataset is unavailable.

Ohkawa_AssemblyHands_Towards_Egocentric_Activity_Understanding_via_3D_Hand_Pose_Estimation_CVPR_2023

Abstract We present AssemblyHands , a large-scale benchmark dataset with accurate 3D hand pose annotations, to facili- tate the study of egocentric activities with challenging hand- object interactions. The dataset includes synchronized ego- centric and exocentric images sampled from the recent As- sembly101 dataset, in which participants assemble and dis- assemble take-apart toys. To obtain high-quality 3D hand pose annotations for the egocentric images, we develop an efficient pipeline, where we use an initial set of manual an- notations to train a model to automatically annotate a much larger dataset. Our annotation model uses multi-view fea- ture fusion and an iterative refinement scheme, and achieves an average keypoint error of 4.20 mm, which is 85% lower than the error of the original annotations in Assembly101. AssemblyHands provides 3.0M annotated images, includ- ing 490K egocentric images, making it the largest existing benchmark dataset for egocentric 3D hand pose estimation. Using this data, we develop a strong single-view baseline of 3D hand pose estimation from egocentric images. Further- more, we design a novel action classification task to evalu- ate predicted 3D hand poses. Our study shows that having higher-quality hand poses directly improves the ability to recognize actions.

1. Introduction Recognizing human activities is a decades-old problem in computer vision [17]. With recent advancements in user- assistive augmented reality and virtual reality (AR/VR) sys- tems, there is an increasing demand for recognizing ac- tions from the egocentric (first-person) viewpoint. Popu- lar AR/VR headsets such as Microsoft HoloLens, Magic Leap, and Meta Quest are typically equipped with egocen- tric cameras to capture a user’s interactions with the real or virtual world. In these scenarios, the user’s hands manip- * Work done during internship. Action classification OriginalAssembly101AssemblyHands(Ours)û Exocentric images3D pose annotations + egocentric images TrainingMulti-view annotationEgocentric image3D hand pose“position”ûPose estimationAction classification Egocentric image3D hand pose“screw”üü Pose estimationFigure 1. High-quality 3D hand poses as an effective represen- tation for egocentric activity understanding. AssemblyHands provides high-quality 3D hand pose annotations computed from multi-view exocentric images sampled from Assembly101 [28], which originally comes with inaccurate annotations computed from egocentric images (see the incorrect left-hand pose predic- tion). As we experimentally demonstrate on an action classifica- tion task, models trained on high-quality annotations achieve sig- nificantly higher accuracy. ulating objects is a very important modality of interaction. In particular, hand poses ( e.g., 3D joint locations) play a central role in understanding and enabling hand-object in- teraction [3, 18], pose-based action recognition [7, 20, 28], and interactive interfaces [10, 11]. Recently, several large-scale datasets for understanding egocentric activities have been proposed, such as EPIC- KITCHENS [5], Ego4D [8], and Assembly101 [28]. In par- ticular, Assembly101 highlights the importance of 3D hand poses in recognizing procedural activities such as assem- bling toys. 3D hand poses are compact representations, and are highly indicative of actions and even the objects that are interacted with– for example, the “screwing” hand motion is 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. 12999 Step1: Auto-annotation from exocentric images Step2: 3D hand pose estimation from egocentric images MVExoNetIterative refinement 3D hand joints (GT) SVEgoNet Training Step3: Evaluation by action classification Action classifier time verb: “position” Figure 2. Construction of AssemblyHands dataset and a benchmark task for egocentric 3D hand pose estimation. We first use manual annotations and an automatic annotation network (MVExoNet) to generate accurate 3D hand poses for multi-view images sampled from the Assembly101 dataset [28]. These annotations are used to train a single-view 3D hand pose estimation network (SVEgoNet) from egocentric images. Finally, the predicted hand poses are evaluated by the action classification task. a strong cue for the presence of a screwdriver. Notably, the authors of Assembly101 found that, for classifying assem- bly actions, learning from 3D hand poses is more effective than solely using video features. However, a drawback of this study is that the 3D hand pose annotations in Assem- bly101 are not always accurate, as they are computed from an off-the-shelf egocentric hand tracker [11]. We observed that the provided poses are often inaccurate (see Fig. 1), es- pecially when hands are occluded by objects from the ego- centric perspective. Thus, the prior work has left us with an unresolved question: How does the quality of 3D hand poses affect action recognition performance? To systematically answer this question, we propose a new benchmark dataset named AssemblyHands . It in- cludes a total of 3.0M images sampled from Assembly101, annotated with high-quality 3D hand poses. We not only acquire manual annotations, but also use them to train an accurate automatic annotation model that uses multi-view feature fusion from exocentric ( i.e., third-person) images; please see Fig. 2 for an illustration. Our model achieves 4.20 mm average keypoint error compared to manual anno- tations, which is 85% lower than the original annotations provided in Assembly101. This automatic pipeline enables us to efficiently scale annotations to 490K egocentric im- ages from 34 subjects, making AssemblyHands the largest egocentric hand pose dataset to date, both in terms of scale and subject diversity. Compared to recent hand-object inter- action datasets, such as DexYCB [3] and H2O [18], our As- semblyHands features significantly more hand-object com- binations, as each multi-part toy can be disassembled and assembled at will, Given the annotated dataset, we first develop a strong baseline for egocentric 3D hand pose estimation, using 2.5D heatmap optimization and hand identity classification.Then, to evaluate the effectiveness of predicted hand poses, we propose a novel evaluation scheme: action classification from hand poses. Unlike prior benchmarks on egocentric hand pose estimation [7, 18, 24], we offer detailed analysis of the quality of 3D hand pose annotation, its influence on the performance of an egocentric pose estimator, and the utility of predicted poses for action classification. Our contributions are summarized as follows: • We offer a large-scale benchmark dataset, dubbed AssemblyHands, with 3D hand pose annotations for 3.0M images sampled from the Assembly101 dataset, including 490K egocentric images. • We propose an automatic annotation pipeline with multi-view feature fusion and iterative refinement, leading to 85% error reduction in the hand pose an- notations. • We define a benchmark task for egocentric 3D hand pose estimation with the evaluation from action classi- fication. We provide a strong single-view baseline that optimizes 2.5D keypoint heatmaps and classifies hand identity. Our results confirm that having high-quality 3D hand poses significantly improves egocentric ac- tion recognition performance.

Mao_Doubly_Right_Object_Recognition_A_Why_Prompt_for_Visual_Rationales_CVPR_2023

Abstract Many visual recognition models are evaluated only on their classification accuracy, a metric for which they obtain strong performance. In this paper, we investigate whether computer vision models can also provide correct rationales for their predictions. We propose a “doubly right” ob- ject recognition benchmark, where the metric requires the model to simultaneously produce both the right labels as well as the right rationales. We find that state-of-the-art visual models, such as CLIP , often provide incorrect ratio- nales for their categorical predictions. However, by trans- ferring the rationales from language models into visual rep- resentations through a tailored dataset, we show that we can learn a “why prompt, ” which adapts large visual repre- sentations to produce correct rationales. Visualizations and empirical experiments show that our prompts significantly improve performance on doubly right object recognition, in addition to zero-shot transfer to unseen tasks and datasets.

1. Introduction Computer vision models today are able to achieve high accuracy – sometimes super-human – at correctly recogniz- ing objects in images. However, most models today are not evaluated on whether they get the prediction right for the right reasons [14, 19, 48, 53]. Learning models that can ex- plain their own decision is important for building trustwor- thy systems, especially in applications that require human- machine interactions [2, 15, 37, 50]. Rationales that justify the prediction can largely improve user trust [54], which is a crucial metric that the visual recognition field should push forward in the future. Existing methods in interpretability have investigated how to understand which features contribute to the mod- els’ prediction [33,34,44,46,47,52,60]. However, saliency explanations are often imprecise, require domain expertise to understand, and also cannot be evaluated. [20, 22] have instead explored verbal rationales to justify the decision- making. However, they require manual collections of the Vision ModelGreek Salad Vision Model Why Prompt This is a photo of a greek salad because there are various toppings such as diced tomatoes, onions, peppers, and jalapenosFigure 1. Visual reasoning for doubly right object recognition task. Motivated by prompting in NLP, we learn a whyprompt from mul- timodal data, which allows us to instruct visual models to predict both the right category and the correct rationales that justify the prediction. plausible rationales in the first place, which subsequently are limited to small-scale datasets and tasks [24, 56]. Scalable methods for explainability have been developed in natural language processing (NLP) through prompting . By adding additional instructions to the input, such as the sentence “think step-by-step,” language models then output descriptions of their reasoning through the chain of thought process [57]. Since the explanations are verbal, they are easily understandable by people, and since the mechanism emerges without explicit supervision, it is highly scalable. In this paper, we investigate whether visual representations can also explain their reasoning through visual chain-of- thought prompts. Our paper first introduces a benchmark for doubly right object recognition, where computer vision models must pre- dict both correct categorical labels as well as correct ra- tionales. Our benchmark is large, and covers many cate- gories and datasets. We found that the visual representa- tions do not have double right capability out-of-the-box on our benchmark. The recent large-scale image-language pre- trained models [41, 49] can retrieve open-world language descriptions that are closest to the image embedding in the feature space, serving as verbal explanations. However, the models often select the wrong rationales. Instead, we propose a framework to explicitly trans- fer the chain-of-thought reasoning from NLP models into 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. 2722 vision models. We first query the large-scale language model [9] via the chain-of-thought reasoning for object cat- egory, where we obtain language rationales that explain dis- criminative features for an object. We then collect images containing both the category and the rationale features us- ing Google image search. We then train visual prompts to transfer the verbal chain of thought to visual chain of thought with contrastive learning, where features of im- ages and their rationales are pulled together. Our “why” prompts obtain up to 26 points gain at doubly right per- formance when evaluated on our benchmark. In addition, visualizations and quantitative results show that our why prompts zero-shot transfer to unseen tasks and datasets. We believe this “doubly right” object recognition task is a fu- ture direction which the visual recognition field should go forward for. Our data and code is available at https: //github.com/cvlab-columbia/DoubleRight .

Pang_Standing_Between_Past_and_Future_Spatio-Temporal_Modeling_for_Multi-Camera_3D_CVPR_2023

Abstract This work proposes an end-to-end multi-camera 3D multi-object tracking (MOT) framework. It emphasizes spatio-temporal continuity and integrates both past and fu- ture reasoning for tracked objects. Thus, we name it “Past- and-Future reasoning for Tracking” (PF-Track). Specifi- cally, our method adopts the “tracking by attention” frame- work and represents tracked instances coherently over time with object queries. To explicitly use historical cues, our “Past Reasoning” module learns to refine the tracks and enhance the object features by cross-attending to queries from previous frames and other objects. The “Future Rea- soning” module digests historical information and predicts robust future trajectories. In the case of long-term occlu- sions, our method maintains the object positions and en- ables re-association by integrating motion predictions. On the nuScenes dataset, our method improves AMOTA by a large margin and remarkably reduces ID-Switches by 90% compared to prior approaches, which is an order of mag- nitude less. The code and models are made available at https://github.com/TRI-ML/PF-Track.

1. Introduction Reasoning about object trajectories in 3D is the cor- nerstone of autonomous navigation. While many LiDAR- based approaches exist [36, 58, 63], their applicability is limited by the cost and reliability of the sensor. Detecting, tracking, and forecasting object trajectories only with cam- eras is hence a critical problem. Significant progress has been achieved on these tasks separately, but they have been historically primarily studied in isolation and combined into a full-stack pipeline in an ad-hoc fashion. In particular, 3D detection has attracted a lot of atten- tion [20,24,25,28,53], but associating these detections over time has been mostly done independently from localiza- *Work done while interning at Toyota Research Institute. †Corresponding to Ziqi Pang at ziqip2@illinois.edu and Yu- Xiong Wang at yxw@illinois.edu . t=0Front-rightCameraBack-rightCameraBack-rightCameraBackCamerat=t!t=t"t=T…… PastReasoningàBetterTrackQualityFutureReasoningàAddressOcclusions…………Figure 1. We visualize the output of our model by projecting pre- dicted 3D bounding boxes onto images. In the beginning, image- based detection can be inaccurate ( t= 0) due to depth ambiguity. With “Past Reasoning,” the bounding box quality ( t=t1) gradu- ally improves by leveraging historical information. With “Future Reasoning,” our PF-Track predicts the long-term motions of ob- jects and maintains their states even under occlusions ( t=t2) and camera switches. This enables re-association without explicit re- identification ( t=T), as the object ID does not switch. Our PF- Track further combines past and future reasoning in a joint frame- work to improve spatio-temporal coherence. tion [19, 31, 43]. Recently, a few approaches to end-to- end detection and tracking have been proposed, but they operate on neighboring frames and fail to integrate longer- term spatio-temporal cues [7, 12, 33, 65]. In the predic- tion literature, on the other hand, it is common to assume the availability of ground truth object trajectories and HD- Maps [4,8,11,59]. A few attempts for a more realistic eval- uation have been made [16, 21], focusing only on the pre- diction performance. In this paper, we argue that multi-object tracking can be dramatically improved by jointly optimizing the detection- tracking-prediction pipeline, especially in a camera-based system. We provide an intuitive example from our real- world experiment in Fig. 1. At first, the pedestrian is fully visible, but a model with only single-frame information makes a prediction with large deviation (frame t= 0 in Fig. 1). After this, integrating the temporal information from the past gradually corrects the error over time (frame t=t1in Fig. 1), by capitalizing on the notion of spatio- temporal continuity. Moreover, as the pedestrian becomes fully occluded (frame t=t2in Fig. 1), we can still pre- dict their location by using the aggregated past informa- 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. 17928 tion to estimate a future trajectory. Finally, we can suc- cessfully track the pedestrian on re-appearance even on a different camera via long-term prediction, resulting in cor- rect re-association (frame t=Tin Fig. 1). The above ro- bust spatio-temporal reasoning is enabled by seamless, bi- directional integration of past and future information, which starkly contrasts with the mainstream pipelines for vision- based, multi-camera, 3D multi-object tracking (3D MOT). To this end, we propose an end-to-end framework for joint 3D object detection, tracking, and trajectory predic- tion for the task of 3D MOT, as shown in Fig. 2, adopting the “tracking by attention” [34,64,65] paradigm. Compared to our closest baseline under the same paradigm [65], we are different in explicit past and future reasoning: a 3D object query consistently represents the object over time, propa- gates the spatio-temporal information of the object across frames, and generates the corresponding bounding boxes and future trajectories. To exploit spatio-temporal cues, our algorithm leverages simple attention operations to capture object dynamics and interactions, which are then used for track refinement and robust, long-term trajectory prediction. Finally, we close the loop by integrating predicted trajecto- ries back into the tracking module to replace missing detec- tions ( e.g., due to an occlusion). To highlight the capabil- ity of joint past and future reasoning, our method is named “Past-and-Future reasoning for Tracking” (PF-Track). We provide a comprehensive evaluation of PF-Track on nuScenes [4] and demonstrate that joint modeling of past and future information provides clear benefits for object tracking. In particular, PF-Track decreases ID-Switches by over 90% compared to previous multi-camera 3D MOT methods. To summarize, our contributions are as follows. 1. We propose an end-to-end vision-only 3D MOT frame- work that utilizes object-level spatio-temporal reasoning for both past and future information. 2. Our framework improves the quality of tracks by cross- attending to features from the “ past.” 3. We propose a joint tracking and prediction pipeline, whose constituent part is “Future Reasoning”, and demonstrate that tracking can explicitly benefit from long-term prediction into the “ future .” 4. Our method establishes new state-of-the-art on large- scale nuScenes dataset [4] with significant improvement for both AMOTA and ID-Switch.

Li_StyleGene_Crossover_and_Mutation_of_Region-Level_Facial_Genes_for_Kinship_CVPR_2023

Abstract High-fidelity kinship face synthesis has many potential applications, such as kinship verification, missing child identification, and social media analysis. However, it is challenging to synthesize high-quality descendant faces with genetic relations due to the lack of large-scale, high- quality annotated kinship data. This paper proposes RFG (Region-level Facial Gene) extraction framework to address this issue. We propose to use IGE (Image-based Gene En- coder), LGE (Latent-based Gene Encoder) and Gene De- coder to learn the RFGs of a given face image, and the relationships between RFGs and the latent space of Style- GAN2. As cycle-like losses are designed to measure the L2 distances between the output of Gene Decoder and image encoder, and that between the output of LGE and IGE, only *Corresponding Authorface images are required to train our framework, i.e. no paired kinship face data is required. Based upon the pro- posed RFGs, a crossover and mutation module is further designed to inherit the facial parts of parents. A Gene Pool has also been used to introduce the variations into the mu- tation of RFGs. The diversity of the faces of descendants can thus be significantly increased. Qualitative, quantita- tive, and subjective experiments on FIW, TSKinFace, and FF-Databases clearly show that the quality and diversity of kinship faces generated by our approach are much better than the existing state-of-the-art methods.

1. Introduction Humans can identify kinship through photographs based on the resemblance between parents and children. Many works have investigated this intrinsic relation in the fields 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. 20960 of kinship verification [9, 32, 42] and genetics [4, 5, 8, 19]. With the popularity of face synthesis and editing technol- ogy in recent years, high-fidelity kinship face synthesis has also attracted much attention. This task, aiming to synthe- size the faces of descendants based on the appearance of the parents, has many potential applications, such as find- ing long-lost children, crime investigations, kinship verifi- cation, and multimedia social applications. In recent years, many efforts have been made to make use of generative models [7, 12, 15, 16, 27, 29, 38, 43, 45, 48] for kinship face synthesis. These works can be catego- rized into two paradigms: one-stage and two-stage. The one-stage paradigm [12, 29, 38, 45] treats this problem as an image-to-image translation task and trains a one-to-one kinship face generator with paired data. However, these ap- proaches can only produce low-resolution images and the resultant images can be blurry and lack diversity. Further, it would be quite difficult to obtain annotated kinship data. By contrast, the two-stage paradigm [7, 15, 16, 27, 43, 48] first extracts the genetic representation and assembles them into children’s representation based on the parents’ faces. To obtain genetic representation, existing methods try to learn the inheritance and variation of facial appearances by training deep neural networks [14, 27, 48] or via a knowl- edge rule [7]. However, the learned genetic representation is prone to overfitting due to the lack of high-quality kin- ship annotated training data, resulting in a lack of diversity in the generated children. In addition, these methods cannot provide fine-grained attributes representation, and thus the generated facial attributes lack interpretability. In this paper, the facial genetic process is abstracted as the exchange and mutation of the parents’ facial parts. We propose an Image-based Gene Encoder (IGE) to construct an independent representation for each facial part, called a Region-level Facial Gene (RFG), which is used to con- trol the synthesis of facial regions. We further simulate the crossover and mutation process to assemble the RFGs of de- scendants by using those of the parents, and our proposed Gene Pool used in the mutation process can significantly increase the diversity of the generated descendants. We use the pre-trained StyleGAN2 [24] as the generator to synthe- size high-fidelity faces. To achieve this, we use a Gene De- coder to map RFGs to the W+space of StyleGAN2. Since IGE requires a facial parsing mask to generate the RFG, we additionally train a Latent-based Gene Encoder (LGE) to directly map the latent code of StyleGAN2 to RFGs. Thus, facial parsing mask is not required for the RFG extraction in the inference stage. The main contributions of this paper are summarized as follows: • We propose StyleGene to synthesize high-fidelity kin- ship faces with controllable facial genetic regions, via modeling the facial genetic relations based on the pro- posed region-level facial genes.• A novel genetic strategy is further introduced by sim- ulating the crossover and mutation process to generate the RFGs of descendants. We introduce a Gene Pool into the mutation process to significantly increase the diversity of the kinship face. • We validate the effectiveness of our approach on sev- eral benchmarks, demonstrating the superiority of our StyleGene framework over other state-of-the-art meth- ods, in terms of the quality and diversity of the gener- ated kinship faces.

Lu_Decomposed_Soft_Prompt_Guided_Fusion_Enhancing_for_Compositional_Zero-Shot_Learning_CVPR_2023

Abstract Compositional Zero-Shot Learning (CZSL) aims to rec- ognize novel concepts formed by known states and objects during training. Existing methods either learn the combined state-object representation, challenging the generalization of unseen compositions, or design two classifiers to identify state and object separately from image features, ignoring the intrinsic relationship between them. To jointly eliminate the above issues and construct a more robust CZSL system, we propose a novel framework termed Decomposed Fusion withSoftPrompt (DFSP)1, by involving vision-language models (VLMs) for unseen composition recognition. Specif- ically, DFSP constructs a vector combination of learnable soft prompts with state and object to establish the joint rep- resentation of them. In addition, a cross-modal decomposed fusion module is designed between the language and image branches, which decomposes state and object among lan- guage features instead of image features. Notably, being fused with the decomposed features, the image features can be more expressive for learning the relationship with states and objects, respectively, to improve the response of un- seen compositions in the pair space, hence narrowing the domain gap between seen and unseen sets. Experimental results on three challenging benchmarks demonstrate that our approach significantly outperforms other state-of-the- art methods by large margins.

1. Introduction Given an unseen concept, such as green tiger , even though this is a nonexistent stuff humans have never seen, they may associate the known state green with an image of tiger immediately. Inspired by this, Compositional Zero- Shot Learning (CZSL) is proposed with the purpose of *Song Guo and Jingcai Guo are the corresponding authors 1Code is available at: https://github.com/Forest-art/ DFSP.git a photo of [state] [object]narrow the domian gaplearn the joint representation old cat (unseen)DFSP vftfofsf Fusion seenunseenolddry catdogs vff o vffFigure 1. The overview of DFSP. Our method aims to narrow the domain gap between seen and unseen compositions by fus- ing decomposed features foandfswith image feature fv, while learn the joint representation between state and object in language branch. Being fused with the state and object features, image fea- ture can learn the response of them respectively and improve the sensitiveness of unseen compositions. equipping models with the ability to recognize novel con- cepts generated as humans do. Specifically, CZSL learns on visible primitive composed concepts (state and object) in the training phase, and recognizes unseen compositions in the inference phase. Some prior algorithms [20, 26] design two classifiers to identify state and object separately, while these mod- els overlook the intrinsic relation between them. After the primitive concepts are obtained, the association between state and object could be established again through graph neural network (GNN) [24] or external knowledge compo- sitions [14]. Nevertheless, these are post-processing meth- ods and these classifiers are separated from image features with strong correlation, ignoring entanglement. Some other methods [28, 29] are to directly treat the combination as an entity, converting CZSL into a general zero-shot recognition problem. Generally, the visual features are projected into a shared semantic space and the distance between entities is optimized, such as Euclidean distance [44]. If too much 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. 23560 attention is paid to the composed concepts in the training stage, the model can not be generalized well to unseen com- positions, causing the domain gap between seen and unseen sets. In summary, these methods are all visual recognition models, which are limited by the strong entanglement of states and objects in image features. In contrast, we focus on designing novel approaches based on vision-language models (VLMs) to cope with CZSL challenges. Since state and object are two separate words in the text, they are less entangled in language fea- tures than image features and could be decomposed more easily and precisely. Certainly, state and object are also intrinsically linked in the text, such as ripe apple instead ofold apple . Constructing the combination in the form of text can also establish the joint representation of state and object to pair with images. Meanwhile, the decomposed state and object features can also be independently associ- ated with the image feature, easing the excessive bias of the model towards seen compositions and enhancing the unseen response (shown in Fig. 1). To improve CZSL with VLMs, we design Decomposed Fusion with SoftPrompt (DFSP), an efficient framework aimed to both learn about the joint representation of primitive concepts and shrink the domain gap between seen and unseen composition sets, as shown in Fig. 2. To be specific, DFSP is designed as a fully learnable soft prompt including prefix, state and object, which con- structs the joint representation between primitive concepts and can be fine-tuned well for new supervised tasks. We then design a decomposed fusion module (DFM) for state and object, which decomposes features extracted from text encoder, such as Bert [6], etc. Meanwhile, the decomposed language features and image features of DFSP interact with information in a cross-modal fusion module, which is cru- cial for learning high-quality language-aware visual repre- sentations. During the phase of fusion, the image can es- tablish separate relationships with the state and object, and then is paired with the composed prompt feature in the pair space, improving its response even for unseen compositions to shrink the domain gap. Generally, this paper makes the following contributions: • A novel framework named Decomposed Fusion with Soft Prompt (DFSP) is proposed, which is based on vision-language paradigm aiming to cope with CZSL. • The Decomposed Fusion Module is designed for CZSL specifically, which decomposes the concepts of language features and fuses them with image features to improve the response of unseen compositions. • We design a learnable soft prompt to construct the joint-representation of state and object, which can be more precisely decomposed than images.• Extensive experiments demonstrate the effectiveness of DFSP, which greatly outperforms the state-of-the- art CZSL approaches on both closed-world and open- world.

Li_SECAD-Net_Self-Supervised_CAD_Reconstruction_by_Learning_Sketch-Extrude_Operations_CVPR_2023

Abstract Reverse engineering CAD models from raw geometry is a classic but strenuous research problem. Previous learning-based methods rely heavily on labels due to the supervised design patterns or reconstruct CAD shapes that are not easily editable. In this work, we introduce SECAD- Net, an end-to-end neural network aimed at reconstructing compact and easy-to-edit CAD models in a self-supervised manner. Drawing inspiration from the modeling language that is most commonly used in modern CAD software, we propose to learn 2D sketches and 3D extrusion parame- ters from raw shapes, from which a set of extrusion cylin- ders can be generated by extruding each sketch from a 2D plane into a 3D body. By incorporating the Boolean op- eration ( i.e., union), these cylinders can be combined to closely approximate the target geometry. We advocate the use of implicit fields for sketch representation, which al- lows for creating CAD variations by interpolating latent codes in the sketch latent space. Extensive experiments on both ABC and Fusion 360 datasets demonstrate the effec- tiveness of our method, and show superiority over state-of- the-art alternatives including the closely related method for supervised CAD reconstruction. We further apply our ap- *Corresponding author: jianwei.guo@nlpr.ia.ac.cnproach to CAD editing and single-view CAD reconstruc- tion. Code will be released at https://github.com/ BunnySoCrazy/SECAD-Net .

1. Introduction CAD reconstruction is one of the most sought-after ge- ometric modeling technologies, which plays a substantial role in reverse engineering in case of the original design document is missing or the CAD model of a real object is not available. It empowers users to reproduce CAD mod- els from other representations and supports the designer to create new variations to facilitate various engineering and manufacturing applications. The advance in 3D scanning technologies has promoted the paradigm shift from time-consuming and laborious manual dimensions to automatic CAD reconstruction. A typical line of works [3,6,35,47] first reconstructs a polygon mesh from the scanned point cloud, then followed by mesh segmentation and primitive extraction to obtain a bound- ary representation (B-rep). Finally, a CAD shape parser is applied to convert the B-rep into a sequence of modeling operations. Recently, inspired by the substantial success of point set learning [1, 32, 49] and deep 3D representa- tions [9,28,30], a number of methods have exploited neural 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. 16816 networks to improve the above pipeline, e.g., detecting and fitting primitives to raw point clouds directly [25,27,40]. A few works ( e.g., CSG-Net [39], UCSG-Net [19], and CSG- Stump [33]) further parse point cloud inputs into a construc- tive solid geometry (CSG) tree by predicting a set of prim- itives that are then combined with Boolean operations. Al- though achieving encouraging compact representation, they only output a set of simple primitives with limited types (e.g., planes, cylinders, spheres), which restricts their rep- resentation capability for reconstructing complex and more general 3D shapes. CAPRI-Net [59] introduces quadric sur- face primitives and the difference operation based on BSP- Net [8] to produce complicated convex and concave shapes via a CSG tree. However, controlling the implicit equation and parameters of quadric primitives is difficult for design- ers to edit the reconstructed models. Thus, the editability of those methods is quite limited. In this paper, we develop a novel and versatile deep neu- ral framework, named SECAD-Net, to reconstruct high- quality and editable CAD models. Our approach is inspired by the observation that a CAD model is usually designed as a command sequence of operations [7, 38, 50, 51, 57], i.e., a set of planar 2D sketches are first drawn then extruded into 3D solid shapes for Boolean operations to create the final model. At the heart of our approach is to learn the sketch and extrude modeling operations, rather than CSG with parametric primitives. To determine the position and axis of each sketch plane, SECAD-Net first learns multiple extrusion boxes to decompose the entire shape into multiple local regions. Afterward, for the local profile in each box, we utilize a fully connected network to learn the implicit representation of the sketch. An extrusion operator is then designed to calculate the implicit expression of the cylinders according to the predicted sketch and extrusion parameters. We finally apply a union operation to assemble all extrusion cylinders into the final CAD model. Benefiting from our representation, our approach is flex- ible and efficient to construct a wide range of 3D shapes. As the predictions of our method are fully interpretable, it allows users to express their ideas to create variations or im- prove the design by operating on 2D sketches or 3D cylin- ders intuitively. To summarize, our work makes the follow- ing contributions: • We present a novel deep neural network for reverse en- gineering CAD models with self-supervision, leading to faithful reconstructions that closely approximate the target geometry. • SECAD-Net is capable of learning implicit sketches and differentiable extrusions from raw 3D shapes with- out the guidance of ground truth sketch labels. • Extensive experiments demonstrate the superiority of SECAD-Net through comprehensive comparisons. Wealso showcase its immediate applications to CAD in- terpolation, editing, and single-view reconstruction.

Luo_Zero-Shot_Model_Diagnosis_CVPR_2023

Abstract When it comes to deploying deep vision models, the be- havior of these systems must be explicable to ensure confi- dence in their reliability and fairness. A common approach to evaluate deep learning models is to build a labeled test set with attributes of interest and assess how well it per- forms. However, creating a balanced test set (i.e., one that is uniformly sampled over all the important traits) is of- ten time-consuming, expensive, and prone to mistakes. The question we try to address is: can we evaluate the sensi- tivity of deep learning models to arbitrary visual attributes without an annotated test set ? This paper argues the case that Zero-shotModel Diag- nosis (ZOOM) is possible without the need for a test set nor labeling. To avoid the need for test sets, our system relies on a generative model and CLIP . The key idea is enabling the user to select a set of prompts (relevant to the prob- lem) and our system will automatically search for seman- tic counterfactual images (i.e., synthesized images that flip the prediction in the case of a binary classifier) using the generative model. We evaluate several visual tasks (classi- fication, key-point detection, and segmentation) in multiple visual domains to demonstrate the viability of our method- ology. Extensive experiments demonstrate that our method is capable of producing counterfactual images and offering sensitivity analysis for model diagnosis without the need for a test set.

1. Introduction Deep learning models inherit data biases, which can be accentuated or downplayed depending on the model’s ar- chitecture and optimization strategy. Deploying a computer vision deep learning model requires extensive testing and evaluation, with a particular focus on features with poten- tially dire social consequences (e.g., non-uniform behav- ior across gender or ethnicity). Given the importance of the problem, it is common to collect and label large-scale datasets to evaluate the behavior of these models across attributes of interest. Unfortunately, collecting these test *Equal contribution. CLIP StyleGAN Generator How would a change in [attribute] affect [my model] 's prediction?User Diagnosis Request Neural Tools Diagnosis Outcomes [attribute] = "green eye" "vertical pupil" "pointed ear" ... [my model] = cat/dog classifier ... Zero-shot counterfactual images Zero-shot sensitivity histogram "green eye""vertical pupil" "pointed ear"...... Figure 1. Given a differentiable deep learning model (e.g., a cat/dog classifier) and user-defined text attributes, how can we de- termine the model’s sensitivity to specific attributes without us- ing labeled test data? Our system generates counterfactual images (bottom right) based on the textual directions provided by the user, while also computing the sensitivity histogram (top right). datasets is extremely time-consuming, error-prone, and ex- pensive. Moreover, a balanced dataset, that is uniformly distributed across all attributes of interest, is also typically impractical to acquire due to its combinatorial nature. Even with careful metric analysis in this test set, no robustness nor fairness can be guaranteed since there can be a mis- match between the real and test distributions [25]. This research will explore model diagnosis without relying on a test set in an effort to democratize model diagnosis and lower the associated cost. Counterfactual explainability as a means of model diag- nosis is drawing the community’s attention [5,20]. Counter- factual images visualize the sensitive factors of an input im- age that can influence the model’s outputs. In other words, counterfactuals answer the question: “How can we modify the input image x(while fixing the ground truth) so that the model prediction would diverge from ytoˆy?”. The param- eterization of such counterfactuals will provide insights into identifying key factors of where the model fails. Unlike ex- isting image-space adversary techniques [4,18], counterfac- tuals provide semantic perturbations that are interpretable by humans. However, existing counterfactual studies 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. 11631 quire the user to either collect uniform test sets [10], anno- tate discovered bias [15], or train a model-specific explana- tion every time the user wants to diagnose a new model [13]. On the other hand, recent advances in Contrastive Language-Image Pretraining (CLIP) [24] can help to over- come the above challenges. CLIP enables text-driven ap- plications that map user text representations to visual man- ifolds for downstream tasks such as avatar generation [7], motion generation [37] or neural rendering [22, 30]. In the domain of image synthesis, StyleCLIP [21] reveals that text-conditioned optimization in the StyleGAN [12] latent space can decompose latent directions for image editing, allowing for the mutation of a specific attribute without dis- turbing others. With such capability, users can freely edit semantic attributes conditioned on text inputs. This paper further explores its use in the scope of model diagnosis. The central concept of the paper is depicted in Fig. 1. Consider a user interested in evaluating which factors con- tribute to the lack of robustness in a cat/dog classifier (target model). By selecting a list of keyword attributes, the user is able to (1) see counterfactual images where semantic vari- ations flip the target model predictions (see the classifier score in the top-right corner of the counterfactual images) and (2) quantify the sensitivity of each attribute for the tar- get model (see sensitivity histogram on the top). Instead of using a test set, we propose using a StyleGAN generator as the picture engine for sampling counterfactual images. CLIP transforms user’s text input, and enables model diag- nosis in an open-vocabulary setting. This is a major advan- tage since there is no need for collecting and annotating im- ages and minimal user expert knowledge. In addition, we are not tied to a particular annotation from datasets (e.g., specific attributes in CelebA [16]). To summarize, our proposed work offers three major im- provements over earlier efforts: • The user requires neither a labeled, balanced test dataset, and minimal expert knowledge in order to evaluate where a model fails (i.e., model diagnosis). In addition, the method provides a sensitivity histogram across the attributes of interest. • When a different target model or a new user-defined attribute space is introduced, it is not necessary to re- train our system, allowing for practical use. • The target model fine-tuned with counterfactual im- ages not only slightly improves the classification per- formance, but also greatly increases the distributional robustness against counterfactual images.

Niu_Visibility_Constrained_Wide-Band_Illumination_Spectrum_Design_for_Seeing-in-the-Dark_CVPR_2023

Abstract Seeing-in-the-dark is one of the most important and challenging computer vision tasks due to its wide appli- cations and extreme complexities of in-the-wild scenar- ios. Existing arts can be mainly divided into two threads: 1) RGB-dependent methods restore information using de- graded RGB inputs only ( e.g., low-light enhancement), 2) RGB-independent methods translate images captured under auxiliary near-infrared (NIR) illuminants into RGB domain (e.g., NIR2RGB translation). The latter is very attractive since it works in complete darkness and the illuminants are visually friendly to naked eyes, but tends to be unstable due to its intrinsic ambiguities. In this paper, we try to robustify NIR2RGB translation by designing the optimal spectrum of auxiliary illumination in the wide-band VIS-NIR range, while keeping visual friendliness. Our core idea is to quan- tify the visibility constraint implied by the human vision sys- tem and incorporate it into the design pipeline. By model- ing the formation process of images in the VIS-NIR range, the optimal multiplexing of a wide range of LEDs is auto- matically designed in a fully differentiable manner, within the feasible region defined by the visibility constraint. We also collect a substantially expanded VIS-NIR hyperspec- tral image dataset for experiments by using a customized 50-band filter wheel. Experimental results show that the task can be significantly improved by using the optimized wide-band illumination than using NIR only. Codes Avail- able: https://github.com/MyNiuuu/VCSD .

1. Introduction Seeing-in-the-dark is critical for modern industries, be- cause of its promising applications in nighttime photogra- phy and visual surveillance. However, it remains challeng- ing due to complex degradation mechanisms and dynamics of in-the-wild environments. To achieve this task, a number of methods have been pro- *Corresponding authorposed, which can be roughly divided into two threads. The first thread features RGB-dependent methods [3, 4, 29, 42, 44,45,52] that aim to fully exploit the RGB input, even with severe degradations. These methods have gained great suc- cess through directly learning the mapping from low-light input to normal-light output, in the presence of complex noises and color discrepancies. However, even state-of-the- art methods along this thread may struggle with in-the-wild data captured under nearly complete darkness. In contrast, the second thread features RGB-independent methods [24, 28, 33, 38, 40] for non-interfering surveillance that try to recover RGB information from images of invis- ible ranges, without requiring any RGB input. The most attractive characteristics lie in its applicability to complete darkness and the visual friendliness of auxiliary illumina- tion to naked eyes. NIR2RGB is one of the representative tasks of this thread, which aims to translate near-infrared images to RGB images. As for auxiliary illumination in the NIR range, the in- dustry practice is to use NIR LEDs, usually centered at 850 nm or940 nm . However, the captured images are almost monochromatic and lack visual color and texture, which makes NIR2RGB translation ambiguous. The funda- mental reasons for the ambiguities are two folds: 1) The spectral sensitivities of commodity RGB cameras almost overlap around both 850 nm and940 nm , making it hard to recover three-channel color from a single intensity ob- servation. 2) Reflectance spectra of many materials become almost indistinguishable beyond 850 nm , which leads to ob- vious structure gaps from RGB images. As a result, exist- ing studies that tried to directly convert such NIR images to VIS images, even with the most advanced deep learn- ing techniques, can hardly provide satisfying results due to these fundamental restrictions. In [24], Liu et al. pro- posed to properly multiplex different NIR LEDs, ranging from 700 nm to1000 nm , to robustify the NIR2RGB task, and achieves apparently better results than using traditional 850 nm or940 nm LEDs. However, structure gaps still ex- ist due to the restriction of wavelengths in the NIR range, making the results far from satisfying. 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. 13976 The basic motivation of these methods arises from the in- visibility of human naked eyes to NIR lights, so as to reduce visual interference and light pollution. However, up to now, none of these works have explicitly formulated the visibility of certain illumination. Liu et al. [24] empirically picked up the NIR range beyond 700 nm , and there is a clear tendency that LEDs closer to this prescribed boundary are preferred according to their results. A natural question is: Is there an exact boundary between visible and invisible? This is important since it determines how much information in the VIS range can be utilized to help RGB recovery. Inspired by the aforementioned methods, we propose to quantify and incorporate the human vision system into our model, which enables us to significantly robustify this task via illumination spectrum design in the wide-band spectral range from 420 nm to890 nm . Similar to [24], we directly optimize the spectral curve by training an image enhance- ment model on hyperspectral datasets. Specifically, based on the human vision system, we establish a Visibility Con- strained Spectrum Design (VCSD) model to quantify the visibility of certain spectra, and to assure the prescribed vis- ibility level will not be violated. To achieve this, a visibility threshold ˆΨis introduced, which serves as the visibility up- per bound during the spectrum design process. In practice, this threshold can be changed according to the desired level of visibility, without destroying the validity of our method. According to the upper bounded visibility level, the model scales down the designed LED spectrum (if necessary) to assure that the new spectrum is friendly to naked eyes. Af- ter that, we design a physic-based Imaging Process Simu- lation (IPS) model which synthesizes images using the cor- responding LED spectrum, camera spectral sensitivity, and the reflectance spectrum of the scene. The IPS model also contains a noise model to consider the noise effect during the realistic imaging process. Since we consider the spec- trum from 420 nm to890 nm , we synthesize one VIS image with lights shorter than 700 nm and one VIS-NIR image with the full spectrum. Through deep learning, we directly minimize the reconstruction loss and finally get the optimal LED spectral curve that can be physically realized by driv- ing LEDs with appropriate voltage and current. We evaluate the effectiveness of our model and designed curve on hyperspectral datasets including our proposed and previous [32] datasets. Compared to existing methods, our model clearly achieves superior results, demonstrating the powerfulness of wide-band illumination spectrum design under visibility constraints. The main highlights of this work are: • For the first time, we propose a paradigm that quan- tifies and incorporates the human vision system for seeing-in-the-dark, which enables us to significantly improve the task via illumination spectrum design in a wide-band coverage from 420 nm to890 nm .• A novel Visibility Constrained Spectrum Design (VCSD) model is proposed to formulate and assure the visibility level of certain spectra to human naked eyes during the optimization process. The visibility thresh- old can be changed according to the desired level of visibility, without destroying the validity of the model. • We design a physic-based Imaging Process Simula- tion (IPS) module which synthesizes the input images based on the imaging process and the noise model. • We contribute a VIS-NIR wide-band hyperspectral im- age dataset to supplement existing ones in terms of quality and quantity.

Oh_Recovering_3D_Hand_Mesh_Sequence_From_a_Single_Blurry_Image_CVPR_2023

Abstract Hands, one of the most dynamic parts of our body, suffer from blur due to their active movements. However, previous 3D hand mesh recovery methods have mainly focused on sharp hand images rather than considering blur due to the absence of datasets providing blurry hand images. We first present a novel dataset BlurHand, which contains blurry hand images with 3D groundtruths. The BlurHand is con- structed by synthesizing motion blur from sequential sharp hand images, imitating realistic and natural motion blurs. In addition to the new dataset, we propose BlurHandNet, a baseline network for accurate 3D hand mesh recovery from a blurry hand image. Our BlurHandNet unfolds a blurry input image to a 3D hand mesh sequence to utilize tem- poral information in the blurry input image, while previ- ous works output a static single hand mesh. We demon- strate the usefulness of BlurHand for the 3D hand mesh recovery from blurry images in our experiments. The pro- posed BlurHandNet produces much more robust results on blurry images while generalizing well to in-the-wild images. The training codes and BlurHand dataset are available at https://github.com/JaehaKim97/BlurHand_RELEASE.

1. Introduction Since hand images frequently contain blur when hands are moving, developing a blur-robust 3D hand mesh esti- mation framework is necessary. As blur makes the bound- ary unclear and hard to recognize, it significantly degrades the performance of 3D hand mesh estimation and makes the task challenging. Despite promising results of 3D hand mesh estimation from a single sharp image [5,13,16,17,22], research on blurry hands is barely conducted. A primary reason for such lack of consideration is the ab- sence of datasets that consist of blurry hand images with ac- curate 3D groundtruth (GT). Capturing blurry hand datasets *Authors contributed equally. (a)Examples of the presented BlurHand dataset. (b)Illustration of the temporal unfolding. Figure 1. Proposed BlurHand dataset and BlurHandNet. (a) We present a novel BlurHand dataset, providing natural blurry hand images with accurate 3D annotations. (b) While most pre- vious methods produce a single 3D hand mesh from a sharp im- age, our BlurHandNet unfolds the blurry hand image into three sequential hand meshes. is greatly challenging. The standard way of capturing mark- erless 3D hand datasets [8,23,50] consists of two stages: 1) obtaining multi-view 2D grounds ( e.g., 2D joint coordinates and mask) manually [50] or using estimators [14, 15, 44] and 2) triangulating the multi-view 2D grounds to the 3D space. Here, manual annotations or estimators in the first stage are performed from images. Hence, they become un- reliable when the input image is blurry, which results in tri- angulation failure in the second stage. Contemplating these limitations, we present the Blur- Hand, whose examples are shown in Figure 1a. Our Blur- Hand, the first blurry hand dataset, is synthesized from In- terHand2.6M [23], which is a widely adopted video-based 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. 554 hand dataset with accurate 3D annotations. Following state- of-the-art blur synthesis literature [25, 26, 39], we approxi- mate the blurry images by averaging the sequence of sharp hand frames. As such technique requires high frame rates of videos, we employ a widely used video interpolation method [27] to complement the low frame rate (30 frames per second) of InterHand2.6M. We note that our synthetic blur dataset contains realistic and challenging blurry hands. For a given blurry hand image, the most straightforward baseline is sequentially applying state-of-the-art deblurring methods [3, 29, 30, 46] on blurry images and 3D hand mesh estimation networks [21, 22, 38] on the deblurred image. However, such a simple baseline suffers from two limita- tions. First, since hands contain challenging blur caused by complex articulations, even state-of-the-art deblurring methods could not completely deblur the image. Therefore, the performance of the following 3D hand mesh estima- tion networks severely drops due to remaining blur artifacts. Second, since conventional deblurring approaches only re- store the sharp images corresponding to the middle of the motion, it limits the chance to make use of temporal infor- mation, which might be useful for 3D mesh estimation. In other words, the deblurring process restricts networks from exploiting the motion information in blurry hand images. To overcome the limitations, we propose BlurHandNet, which recovers a 3D hand mesh sequence from a single blurry image, as shown in Figure 1b. Our BlurHandNet effectively incorporates useful temporal information from the blurry hand. The main components of BlurHandNet are Unfolder and a kinematic temporal Transformer (KT- Former). Unfolder outputs hand features of three timesteps, i.e., middle and both ends of the motion [12, 28, 32, 36]. The Unfolder brings benefits to our method in two aspects. First, Unfolder enables the proposed BlurHandNet to out- put not only 3D mesh in the middle of the motion but also 3D meshes at both ends of the motion, providing more in- formative results related to motion. We note that this prop- erty is especially beneficial for the hands, where the motion has high practical value in various hand-related works. For example, understanding hand motion is essential in the do- main of sign language [2,34] and hand gestures [40], where the movement itself represents meaning. Second, extract- ing features from multiple time steps enables the following modules to employ temporal information effectively. Since hand features in each time step are highly correlated, ex- ploiting temporal information benefits reconstructing more accurate 3D hand mesh estimation. To effectively incorporate temporal hand features from the Unfolder, we propose KTFormer as the following mod- ule. The KTFormer takes temporal hand features as input and leverages self-attention to enhance the temporal hand features. The KTFormer enables the proposed BlurHand- Net to implicitly consider both the kinematic structure andtemporal relationship between the hands in three timesteps. The KTFormer brings significant performance gain when coupled with Unfolder, demonstrating that employing tem- poral information plays a key role in accurate 3D hand mesh estimation from blurry hand images. With a combination of BlurHand and BlurHandNet, we first tackle 3D hand mesh recovery from blurry hand im- ages. We show that BlurHandNet produces robust results from blurry hands and further demonstrate that BlurHand- Net generalizes well on in-the-wild blurry hand images by taking advantage of effective temporal modules and Blur- Hand. As this problem is barely studied, we hope our work could provide useful insights into the following works. We summarize our contributions as follows: • We present a novel blurry hand dataset, BlurHand, which contains natural blurry hand images with accu- rate 3D GTs. • We propose the BlurHandNet for accurate 3D hand mesh estimation from blurry hand images with novel temporal modules, Unfolder and KTFormer. • We experimentally demonstrate that the proposed BlurHandNet achieves superior 3D hand mesh estima- tion performance on blurry hands.

Qiu_Looking_Through_the_Glass_Neural_Surface_Reconstruction_Against_High_Specular_CVPR_2023

Abstract Neural implicit methods have achieved high-quality 3D object surfaces under slight specular highlights. However, high specular reflections (HSR) often appear in front of tar- get objects when we capture them through glasses. The complex ambiguity in these scenes violates the multi-view consistency, then makes it challenging for recent methods to reconstruct target objects correctly. To remedy this is- sue, we present a novel surface reconstruction framework, NeuS-HSR, based on implicit neural rendering. In NeuS- HSR, the object surface is parameterized as an implicit signed distance function (SDF). To reduce the interference of HSR, we propose decomposing the rendered image into two appearances: the target object and the auxiliary plane. We design a novel auxiliary plane module by combining physical assumptions and neural networks to generate the auxiliary plane appearance. Extensive experiments on syn- thetic and real-world datasets demonstrate that NeuS-HSR outperforms state-of-the-art approaches for accurate and robust target surface reconstruction against HSR. Code is available at https://github.com/JiaxiongQ/ NeuS-HSR .

1. Introduction Reconstructing 3D object surfaces from multi-view im- ages is a challenging task in computer vision and graph- ics. Recently, NeuS [45] combines the surface render- ing [3, 12, 35, 52] and volume rendering [8, 29], for recon- structing objects with thin structures and achieves good per- formance on the input with slight specular reflections. How- ever, when processing the scenes under high specular reflec- tions (HSR), NeuS fails to recover the target object surfaces, as shown in the second row of Fig. 1. High specular reflec- tions are ubiquitous when we use a camera to capture the target object through glasses. As shown in the first row of Fig. 1, in the captured views with HSR, we can recognize the virtual image in front of the target object. The virtual *Bo Ren is the corresponding author. View 1 View 27 View 56 … …Supervision NeuS NeuS -HSRFigure 1. 3D object surface reconstruction under high specular reflections (HSR). Top: A real-world scene captured by a mobile phone. Middle: The state-of-the-art method NeuS [45] fails to re- construct the target object ( i.e., the Buddha). Bottom: We propose NeuS-HSR, which recovers a more accurate target object surface than NeuS. image introduces the photometric variation on the object surface visually, which degrades the multi-view consistency and encodes extreme ambiguities for rendering, then con- fuses NeuS to reconstruct the reflected objects instead of the target object. To adapt to the HSR scenes, one intuitive solution is firstly applying reflection removal methods to reduce HSR, then reconstructing the target object with the enhanced tar- get object appearance as the supervision. However, most recent single-image reflection removal works [4, 9, 23, 24, 26, 40] need the ground-truth background or reflection as supervision, which is hard to be acquired. Furthermore, for these reflection removal methods, testing scenes should be present in the training sets, which limits their generaliza- tion. These facts demonstrate that explicitly using the re- flection removal methods to enhance the target object ap- pearance is impractical. A recent unsupervised reflection removal approach, NeRFReN [18] decomposes the ren- dered image into reflected and transmitted parts by implicit representations. However, it is limited by constrained view directions and simple planar reflectors. When we apply it to scenes for multi-view reconstruction, as Fig. 3 presents, it takes the target object as the content in the reflected image 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. 20823 Supervision=+rayrayCameraCameraNeuSNeuS-HSR Target ObjectAuxiliary PlaneObjectsurfaceHighspecularreflectionsObjectrenderingweightsPlanerenderingweightsFigure 2. NeuS-HSR. High specular reflections (HSR) make NeuS tend to reconstruct the reflected object in HSR. NeuS-HSR phys- ically decomposes the rendered image into the target object and auxiliary plane parts, which encourages NeuS to focus on the tar- get object. and fails to generate the correct transmitted image for target object recovery. The two-stage intuitive solution struggles in our task as discussed above. To tackle this issue, we consider a more effective decomposition strategy than NeRFReN, to enhance the target object appearance for accurate surface reconstruction in one stage. To achieve our goal, we con- struct the following assumptions: Assumption 1 A scene that suffers from HSR can be de- composed into the target object and planar reflector com- ponents. Except for the target object, HSR and most other contents in a view are reflected and transmitted through the planar reflectors ( i.e., glasses). Assumption 2 Planar reflectors intersect with the camera view direction since all view direction vectors generally point to the target object and pass through planar reflec- tors. Based on the above physical assumptions, we propose NeuS-HSR, a novel object reconstruction framework to re- cover the target object surface against HSR from a set of RGB images. For Assumption 1, as Fig. 2 shows, we de- sign an auxiliary plane to represent the planar reflector since we aim to enhance the target object appearance through it. With the aid of the auxiliary plane, we faithfully separate the target object and auxiliary plane parts from the super- vision. For the target object part, we follow NeuS [45] to generate the target object appearance. For the auxiliary plane part, we design an auxiliary plane module with the view direction as the input for Assumption 2, by utilizing neural networks to generate attributes (including the nor- mal and position) of the view-dependent auxiliary plane. When the auxiliary plane is determined, we acquire the aux- iliary plane appearance based on the reflection transforma- tion [16] and neural networks. Finally, we add two appear- ances and then obtain the rendered image, which is super- vised by the captured image for one-stage training. We conduct a series of experiments to evaluate NeuS- HSR. The experiments demonstrate that NeuS-HSR is su- perior to other state-of-the-art methods on the synthetic dataset and recovers high-quality target objects from HSR- effect images in real-world scenes. Transmitted Image Supervision Reflected Image NeuSFigure 3. Decomposition of NeRFReN [18]. NeRFReN fails to separate specular reflections and the target object appearance in this view, then makes NeuS fail to recover the target object surface. To summarize, our main contributions are as follows: • We propose to recover the target object surface, which suffers from HSR, by separating the target object and aux- iliary plane parts of the scene. • We design an auxiliary plane module to generate the ap- pearance of the auxiliary plane part physically to enhance the appearance of the target object part. • Extensive experiments on synthetic and real-world scenes demonstrate that our method reconstructs more accurate target objects than other state-of-the-art methods quanti- tatively and qualitatively.

Memmel_Modality-Invariant_Visual_Odometry_for_Embodied_Vision_CVPR_2023

Abstract Effectively localizing an agent in a realistic, noisy setting is crucial for many embodied vision tasks. Visual Odome- try (VO) is a practical substitute for unreliable GPS and compass sensors, especially in indoor environments. While SLAM-based methods show a solid performance without large data requirements, they are less flexible and robust w.r.t. to noise and changes in the sensor suite compared to learning-based approaches. Recent deep VO models, however, limit themselves to a fixed set of input modalities, e.g., RGB and depth, while training on millions of sam- ples. When sensors fail, sensor suites change, or modali- ties are intentionally looped out due to available resources, e.g., power consumption, the models fail catastrophically. Furthermore, training these models from scratch is even more expensive without simulator access or suitable exist- ing models that can be fine-tuned. While such scenarios get mostly ignored in simulation, they commonly hinder a model’s reusability in real-world applications. We propose a Transformer-based modality-invariant VO approach that can deal with diverse or changing sensor suites of naviga- tion agents. Our model outperforms previous methods while training on only a fraction of the data. We hope this method opens the door to a broader range of real-world applica- tions that can benefit from flexible and learned VO models.

1. Introduction Artificial intelligence has found its way into many com- mercial products that provide helpful digital services. To in- crease its impact beyond the digital world, personal robotics and embodied AI aims to put intelligent programs into bod- ies that can move in the real world or interact with it [15]. One of the most fundamental skills embodied agents must learn is to effectively traverse the environment around them, allowing them to move past stationary manipulation tasks and provide services in multiple locations instead [40]. The ability of an agent to locate itself in an environment is vi- tal to navigating it successfully [12, 64]. A common setup is to equip an agent with an RGB-D (RGB andDepth ) *Work done on exchange at EPFL Figure 1. An agent is tasked to navigate to a goal location us- ingRGB-D sensors. Because GPS+Compass are not available, the location is inferred from visual observations only. Neverthe- less, sensors can malfunction, or availability can change during test-time (indicated by ∼), resulting in catastrophic failure of the localization. We train our model to react to such scenarios by ran- domly dropping input modalities. Furthermore, our method can be extended to learn from multiple arbitrary input modalities, e.g., surface normals, point clouds, or internal measurements. camera and a GPS+Compass sensor and teach it to nav- igate to goals in unseen environments [2]. With extended data access through simulators [28, 39, 40, 47, 57], photo- realistic scans of 3D environments [7, 28, 46, 56, 58], and large-scale parallel training, recent approaches reach al- most perfect navigation results in indoor environments [55]. However, these agents fail catastrophically in more real- istic settings with noisy, partially unavailable, or failing RGB-D sensor readings, noisy actuation, or no access to GPS+Compass [6, 64]. Visual Odometry (VO) is one way to close this per- formance gap and localize the agent from only RGB-D observations [2], and deploying such a model has been shown to be especially beneficial when observations are noisy [12, 64]. However, those methods are not robust to any sensory changes at the test-time, such as a sensor fail- ing, underperforming, or being intentionally looped out. In practical applications [43], low-cost hardware can also experience serious bandwidth limitations, causing RGB (3 channels) and Depth (1 channel) to be transferred at dif- ferent rates. Furthermore, mobile edge devices must bal- ance battery usage by switching between passive ( e.g.,RGB) and active ( e.g.,LIDAR ) sensors depending on the specific episode. Attempting to solve this asymmetry by keeping 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. 21549 separate models in memory, relying on active sensors, or using only the highest rate modality is simply infeasible for high-speed and real-world systems. Finally, a changing sen- sor suite represents an extreme case of sensor failure where access to a modality is lost during test-time. These points demonstrate the usefulness of a certain level of modality in- variance in a VO framework. Those scenarios decrease the robustness of SLAM-based approaches [32] and limit the transferability of models trained on RGB-D to systems with only a subset or different sensors. We introduce “optional” modalities as an umbrella term to describe settings where input modalities may be of lim- ited availability at test-time. Figure 1 visualizes a typical indoor navigation pipeline, but introduces uncertainty about modality availability ( i.e. at test-time, only a subset of all modalities might be available). While previous approaches completely neglect such scenarios, we argue that explicitly accounting for “optional” modalities already during train- ingof VO models allows for better reusability on platforms with different sensor suites and trading-off costly or unre- liable sensors during test-time. Recent methods [12, 64] use Convolution Neural Network (ConvNet) architectures that assume a constant channel size of the input, which makes it hard to deal with multiple ”optional” modalities. In contrast, Transformers [51] are much more amenable to variable-sized inputs, facilitating the training of models that can optionally accept one or multiple modalities [4]. Transformers are known to require large amounts of data for training from scratch. Our model’s data requirements are significantly reduced by incorporating various biases: We utilize multi-modal pre-training [4, 17, 30], which not only provides better initializations but also improves perfor- mance when only a subset of modalities are accessible dur- ing test-time [4]. Additionally, we propose a token-based action prior. The action taken by the agent has shown to be beneficial for learning VO [35,64] and primes the model towards the task-relevant image regions. We introduce the Visual Odometry Transformer (VOT), a novel modality-agnostic framework for VO based on the Transformer architecture. Multi-modal pre-training and an action prior drastically reduce the data required to train the architecture. Furthermore, we propose explicit modality- invariance training. By dropping modalities during train- ing, a single VOT matches the performance of separate uni- modal approaches. This allows for traversing different sen- sors during test-time and maintaining performance in the absence of some training modalities. We evaluate our method on point-goal navigation in the Habitat Challenge 2021 [1] and show that VOT outper- forms previous methods [35] with training on only 5% of the data. Beyond this simple demonstration, we stress that our framework is modality-agnostic and not limited to RGB-D input or discrete action spaces and can be adaptedto various modalities, e.g., point clouds, surface normals, gyroscopes, accelerators, compass, etc. To the best of our knowledge, VOT is the first widely applicable modality- invariant Transformer-based VO approach and opens up ex- citing new applications of deep VO in both simulated and real-world applications. We make our code available at github.com/memmelma/VO-Transformer.

Luo_GeoLayoutLM_Geometric_Pre-Training_for_Visual_Information_Extraction_CVPR_2023

Abstract Visual information extraction (VIE) plays an important role in Document Intelligence. Generally, it is divided into two tasks: semantic entity recognition (SER) and rela- tion extraction (RE). Recently, pre-trained models for doc- uments have achieved substantial progress in VIE, partic- ularly in SER. However, most of the existing models learn the geometric representation in an implicit way, which has been found insufficient for the RE task since geometric in- formation is especially crucial for RE. Moreover, we reveal another factor that limits the performance of RE lies in the objective gap between the pre-training phase and the fine- tuning phase for RE. To tackle these issues, we propose in this paper a multi-modal framework, named GeoLay- outLM, for VIE. GeoLayoutLM explicitly models the geo- metric relations in pre-training, which we call geometric pre-training. Geometric pre-training is achieved by three specially designed geometry-related pre-training tasks. Ad- ditionally, novel relation heads, which are pre-trained by the geometric pre-training tasks and fine-tuned for RE, are elaborately designed to enrich and enhance the feature rep- resentation. According to extensive experiments on stan- dard VIE benchmarks, GeoLayoutLM achieves highly com- petitive scores in the SER task and significantly outperforms the previous state-of-the-arts for RE ( e.g., the F1 score of RE on FUNSD is boosted from 80.35% to 89.45%)1.

1. Introduction Visual information extraction (VIE) is a critical part in Document AI [3, 29, 47]. It has attracted more and more at- tention from both the academic and industrial community. VIE involves semantic entity recognition (SER, a.k.a. en- tity labeling) and relation extraction (RE, a.k.a. entity link- ing) from visually-rich documents (VrDs) such as forms and receipts [3, 17, 22, 35, 39, 41, 45, 46]. Recent years have witnessed the great power of pre-trained multi-modal mod- els [1, 7, 8, 12, 15, 20–22, 30, 38, 40, 41, 43] in VIE tasks, *Both authors contributed equally to this work. 1https://github.com/AlibabaResearch/AdvancedLiterateMachinery True PositiveFalse PositiveFalse Negative (a)(b) True PositiveFalse PositiveFalse Negative (a)(b)Figure 1. Incorrect relation predictions by the previous state-of- the-art model LayoutLMv3 [15]. (a) LayoutLMv3 tends to link two entities relying more on their semantics than the geometric layout, i.e., the entity “212-450-4785” is linked to “Fax Number” regardless of their relationship in layout. (b) LayoutLMv3 suc- cessfully predicts the link in the upper half part but misses the link below, although both links are similar in geometric layout. These two examples clearly show the importance of geometric infor- mation in relation extraction (RE) . Precision Recall F1 LayoutLMv3 75.82 85.45 80.35 + geometric constraint 79.87 85.45 82.57 Table 1. The RE performance improvement by introducing a sim- ple geometric restriction (on the FUNSD dataset). especially the SER task. Compared with SER, the RE task, which aims at predicting the relation between semantic en- tities in documents, has not been fully explored and remains a challenging problem [12, 22]. RE is essential to provide additional structural information closer to human compre- hension of the VrDs [45]. It makes the open-layout infor- mation extraction possible, e.g., for open-layout key-value linking and form-like items grouping. It is widely accepted that document layout understand- ing is crucial for VIE [1, 7, 8, 15, 21, 22, 30, 38, 40, 41, 43], especially for RE [12, 22]. The geometric relationships, a specific form for describing document layout, are impor- tant for document layout representations [22, 27, 31]. 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. 7092 previous pre-trained models for VrDs learn layout represen- tations implicitly by adding coordinates into the model in- puts, combining the relative position encoding or supervis- ing by alignment-related pre-training tasks like text-image alignment [15, 30, 43] and masked vision language model- ing [1,7,12,15,21,22,22,40,41,43]. However, it is not guar- anteed that the geometric layout information is well learned in these models. Taking the state-of-the-art model Lay- outLMv3 as an example, we find it would make mistakes in certain relatively simple scenarios, where the geometric relations between entities are not complicated. As shown in Fig. 1, LayoutLMv3 seems to link two entities depending more on the semantics than the geometric layout. This in- dicates that its layout understanding is not sufficiently dis- criminative. To further verify our conjecture, we conduct an experiment by filtering the false positive relations using a simple geometric restriction (the linkings between entities should not point up beyond a certain distance), the precision would increase by a large margin (more than 4 points) while the recall is controlled unchanged, as detailed in Tab. 1. This experiment proves that LayoutLMv3 does not fully exploit the useful geometric relationship information. Be- sides, most existing methods did not directly take the rela- tion modeling into consideration in pre-training. They usu- ally adopt token/segment-level classification or regression, which might underperform on downstream tasks related to relation modeling. Therefore, it is necessary to learn a bet- ter layout representation for document pre-trained models by modeling the geometric relationships between entities explicitly during pre-training. During RE fine-tuning, previous works usually learn a task head like a single linear or bilinear layer [12, 22] from scratch. On the one hand, since the higher-level pair re- lationship features, which are beyond the token or text- segment features in documents, are complex, we argue that a single linear or bilinear layer is not always adequate to make full use of the encoded features for RE. On the other hand, the RE task head initialized randomly is prone to overfitting with limited fine-tuning data. Since the pre- trained backbone has shown tremendous potential [4, 5], why not pre-train the task head in some way simultane- ously? Several works [10, 14, 26] have proved that smaller gapbetween pre-training and fine-tuning leads to better per- formance for downstream tasks. Hence, there is still consid- erable room for the design and usage of the RE task head. Based on the above observations, we establish a multi- modal pre-trained framework (termed as GeoLayoutLM ) for VIE, in which a geometric pre-training strategy is de- signed to explicitly utilize the geometric relationships be- tween text-segments, and elaborately-designed RE heads are introduced to mitigate the gap between pre-training and fine-tuning on the downstream relation extraction task. Specifically, three geometric relations are defined: the re-lation between two text-segments ( GeoPair ), that among multiple text-segment pairs ( GeoMPair ), and that among three text-segments ( GeoTriplet ). Correspondingly, three self-supervised pre-training tasks are proposed. GeoPair re- lation is modeled by the Direction and Distance Modeling (DDM ) task in which GeoLayoutLM needs to tell the di- rection of a directed pair and identify whether a segment is the nearest to another one in the direction. Furthermore, we design a brand-new pre-training objective called Detection ofDirection Exceptions ( DDE ) for GeoMPair, enabling our model to capture the common pattern of directions among segment pairs, enhance the pair feature representation and discover the detached ones. For GeoTriplet, we propose aCollinearity Identification of Triplet ( CIT) task to iden- tify whether three segments are collinear, which takes a step forward to the modeling of multi-segments relations. It is important for non-local layout feature learning especially in form-like documents. Additionally, novel relation heads are proposed to learn better relation features, which are pre- trained by the geometric pre-training tasks to absorb prior knowledge about geometry, thus mitigating the gap between pre-training and fine-tuning. Extensive experiments on five public benchmarks demonstrate the effectiveness of the pro- posed GeoLayoutLM. Our contributions are summarized as follows: 1) This paper introduces three geometric relations in dif- ferent levels and designs three brand-new geometric pre-training tasks correspondingly for learning the ge- ometric layout representation explicitly. To the best of our knowledge, GeoLayoutLM is the first to ex- plore the geometric relations of multi-pair and multi- segments in document pre-training. 2) Novel relation heads are proposed to benefit the re- lation modeling. Besides, the relation heads are pre- trained by the proposed geometric tasks and fine-tuned for RE, thus mitigating the object gap between pre- training and fine-tuning. 3) Experimental results on visual information extraction tasks including key-value linking as relation extrac- tion, entity grouping as relation extraction, and seman- tic entity recognition show that the proposed GeoLay- outLM significantly outperforms previous state-of-the- arts with good interpretability. Moreover, our model has notable advantages in few-shot RE learning.

Luo_VideoFusion_Decomposed_Diffusion_Models_for_High-Quality_Video_Generation_CVPR_2023

Abstract A diffusion probabilistic model (DPM), which constructs a forward diffusion process by gradually adding noise to data points and learns the reverse denoising process to gen- erate new samples, has been shown to handle complex data distribution. Despite its recent success in image synthesis, applying DPMs to video generation is still challenging due to high-dimensional data spaces. Previous methods usually adopt a standard diffusion process, where frames in the same video clip are destroyed with independent noises, ignoring the content redundancy and temporal correlation. This work presents a decomposed diffusion process via resolving the per-frame noise into a base noise that is shared among all frames and a residual noise that varies along the time axis. The denoising pipeline employs two jointly-learned networks to match the noise decomposition accordingly. Experiments on various datasets confirm that our approach, termed as VideoFusion, surpasses both *Work done at Alibaba group. †Corresponding author.GAN-based and diffusion-based alternatives in high-quality video generation. We further show that our decomposed formulation can benefit from pre-trained image diffusion models and well-support text-conditioned video creation.

1. Introduction Diffusion probabilistic models (DPMs) are a class of deep generative models, which consist of : i) a diffusion process that gradually adds noise to data points, and ii) a denoising process that generates new samples via iterative denoising [14, 18]. Recently, DPMs have made awesome achievements in generating high-quality and diverse im- ages [20–22, 25, 27, 36]. Inspired by the success of DPMs on image generation, many researchers are trying to apply a similar idea to video prediction/interpolation [13, 44, 48]. While study about DPMs for video generation is still at an early stage [16] and faces challenges since video data are of higher dimensions and involve complex spatial-temporal correlations. Previous DPM-based video-generation methods usually 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. 10209 adopt a standard diffusion process, where frames in the same video are added with independent noises and the temporal correlations are also gradually destroyed in noised latent variables. Consequently, the video-generation DPM is required to reconstruct coherent frames from independent noise samples in the denoising process. However, it is quite challenging for the denoising network to simultaneously model spatial and temporal correlations. Inspired by the idea that consecutive frames share most of the content, we are motivated to think: would it be easier to generate video frames from noises that also have some parts in common? To this end, we modify the standard diffusion process and propose a decomposed diffusion probabilistic model, termed as VideoFusion, for video generation. During the diffusion process, we resolve the per-frame noise into two parts, namely base noise and residual noise , where the base noise is shared by consecutive frames. In this way, the noised latent variables of different frames will always share a common part, which makes the denoising network easier to reconstruct a coherent video. For intuitive illustration, we use the decoder of DALL-E2 [25] to generate images conditioned on the same latent embedding. As shown in Fig. 2a, if the images are generated from independent noises, their content varies a lot even if they share the same condition. But if the noised latent variables share the same base noise, even an image generator can synthesize roughly correlated sequences (shown in Fig. 2b). Therefore, the burden of the denoising network of video-generation DPM can be largely alleviated. Furthermore, this decomposed formulation brings addi- tional benefits. Firstly, as the base noise is shared by all frames, we can predict it by feeding one frame to a large pretrained image-generation DPM with only one forward pass. In this way, the image priors of the pretrained model could be efficiently shared by all frames and thereby facilitate the learning of video data. Secondly, the base noise is shared by all video frames and is likely to be related to the video content. This property makes it possible for us to better control the content or motions of generated videos. Experiments in Sec. 4.7 show that, with adequate training, VideoFusion tends to relate the base noise with video content and the residual noise to motions (Fig. 1). Extensive experiments show that VideoFusion can achieve state-of-the-art results on different datasets and also well support text-conditioned video creation.

Li_Rethinking_Feature-Based_Knowledge_Distillation_for_Face_Recognition_CVPR_2023

Abstract With the continual expansion of face datasets, feature- based distillation prevails for large-scale face recognition. In this work, we attempt to remove identity supervision in student training, to spare the GPU memory from saving massive class centers. However, this naive removal leads to inferior distillation result. We carefully inspect the perfor- mance degradation from the perspective of intrinsic dimen- sion, and argue that the gap in intrinsic dimension, namely the intrinsic gap, is intimately connected to the infamous capacity gap problem. By constraining the teacher’s search space with reverse distillation, we narrow the intrinsic gap and unleash the potential of feature-only distillation. Re- markably, the proposed reverse distillation creates univer- sally student-friendly teacher that demonstrates outstand- ing student improvement. We further enhance its effective- ness by designing a student proxy to better bridge the intrin- sic gap. As a result, the proposed method surpasses state- of-the-art distillation techniques with identity supervision on various face recognition benchmarks, and the improve- ments are consistent across different teacher-student pairs.

1. Introduction Despite the unceasing emergence of larger and more powerful models for face recognition (FR), industrial de- ployment continues to demand for accurate and light- weight solutions. Among other compression techniques like pruning [27] and quantization [21], knowledge distillation (KD) has been proven to be effective in producing high- performing compact model from well-trained teacher. Un- like classic KD [17] and its variants [14, 24, 43, 44] who distill on logits, most of the existing works on FR distill on features [11, 13] or feature-relations [8, 20, 35]. One key *Equal contribution.†Corresponding author. 4.04.55.05.56.0 Teacher's In.D IR34 IR50 IR70 IR1006062646668MR-all Accuracy/%FI FO ReFO (ours) FO In.DFigure 1. IResNet18 (IR18) is distilled by four different teach- ers. Feature-only distillation (FO) shows performance degrada- tion comparing to feature-based distillation with ID supervision (FI). The proposed method (ReFO) significantly uplifts the perfor- mance of FO distillation. For both FI and FO, the student perfor- mance drops with larger teachers of lower intrinsic dimension. In line plot: student performance (%) on MR-all benchmark [9]. In bar plot: teacher’s intrinsic dimension (In.D). reason is that the massive and still growing number of iden- tities (IDs) in FR datasets, such as the 2 million IDs in Web- Face42M [45], make it too expensive to save extra teacher’s class centers for logits distillation. The ground truth supervision from ID labels, which we call ID supervision, is still retained when training student models for better distillation results. Nonetheless, it is not only non-trivial to find the right balancing weight [15, 33], the obtained class centers are also not needed during infer- ence in an open-set FR problem. This motivates the com- plete removal of class centers in the student training for a number of benefits: 1) speed, the student distillation breaks free from the need of keeping any class center, providing further training speed-up with even lower GPU memory oc- cupancy; 2) access to unlabeled dataset, removing the de- pendency on ID labels conveniently opens the door to the vast quantity of unlabeled or uncleaned face images like WebFace260M [45]; and 3) better focus on feature space, which is what really matters in an open-set problem. Hence, in this work, we are motivated to investigate feature distilla- 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. 20156 tion for face recognition without ID supervision, which we callfeature-only (FO) distillation . The capacity gap problem is widely observed in various KD applications [7, 19, 30, 37], where the student finds it increasingly difficult to learn from more powerful teacher due to larger mismatch in network capacity. In FO distilla- tion, the naive removal of ID supervision degrades student performance with more severe capacity gap problem. As shown in Fig. 1, comparing to the conventional feature dis- tillation with ID supervision (FI distillation), the IResNet18 (IR18) students trained by four other teachers all experience drops in performance when ID supervision is removed. Pertinent works commonly agree that differing model sizes cause the capacity gap issue [7, 20, 30, 40]. Some remedies were proposed to mitigate the problem such as early stopping [7] and training teacher assistants as inter- mediate agents [30]. Liu et al. [26] further proved the im- portance of teacher-student structural compatibility. For a given teacher, their best student from Neural Architecture Search outperformed other candidates of similar model size in the search space. However, recent works like [3, 32] showed that teachers of the same structure, same parameter size and comparable accuracy can also have differing dis- tillation results on the same student. Hence, there must be other factors contributing to the capacity gap problem other than model size and model structure. In this work, we argue that the teacher-student gap in in- trinsic dimension, namely the intrinsic gap , plays a part. The intrinsic dimension [2, 16, 36] of a feature space is the minimum number of variables needed to unambigu- ously describe all points in the feature space. Specifically for a model, lower intrinsic dimension is often associated with better generalization power and better performance for both general classification [2] and face recognition [16]. In Fig. 1, as the teacher gets stronger with lower intrinsic di- mension, we observe a drop in student performance with wider intrinsic gap for both FI distillation and FO distilla- tion. If narrower intrinsic gap is related to better distillation result, can the capacity gap problem be mitigated by closing the intrinsic gap? This sparkles the idea that whether it is possible to narrow the intrinsic gap by raising teacher’s in- trinsic dimension for easier student-learning, neither chang- ing its model size nor model structure. Firstly, we revisit FO distillation and point out the intrin- sic gap as another factor that could cause ineffective dis- tillation. Then a reverse distillation strategy is proposed to solve the problem by injecting knowledge about higher intrinsic dimensional feature space into the teacher train- ing. With reverse-distilled teachers, students trained with just FO distillation loss like mean-square-error (MSE) show performance on par or even better than competitors trained by sophisticatedly designed distillation loss with ID super- vision [20, 35]. The proposed method is thus fast and ver-satile, it can be online or offline and easily portable to unla- beled datasets. On top of that, we further improve the dis- tillation results by allowing the teacher to learn from more light-weight student proxies. This better closes the intrin- sic gap and we are able to obtain state-of-the-art (SOTA) student models on popular face recognition benchmarks. To summarize, the contribution of this work includes: • We reconsider the capacity gap issue in FO distillation and provide an alternative view from the perspective of the intrinsic dimension. The gap in the intrinsic di- mension between the teacher and the student is found to be related to the distillation performance. • We propose a novel training scheme that narrows the teacher-student intrinsic gap via reverse distillation in the teacher training. Furthermore, we enhance its ef- fectiveness by designing light-weight student proxies as the reverse distillation targets. Students trained by the new teachers show consistent performance im- provement on FO distillation. • Our method pushes the limit of FO distillation with easier-to-learn teacher. With only feature distillation loss, resulting students are shown to be superior than students trained by other SOTA distillation techniques with ID supervision.

Nguyen_Re-Thinking_Model_Inversion_Attacks_Against_Deep_Neural_Networks_CVPR_2023

Abstract Model inversion (MI) attacks aim to infer and recon- struct private training data by abusing access to a model. MI attacks have raised concerns about the leaking of sen- sitive information (e.g. private face images used in train- ing a face recognition system). Recently, several algorithms for MI have been proposed to improve the attack perfor- mance. In this work, we revisit MI, study two fundamental issues pertaining to all state-of-the-art (SOTA) MI algo- rithms , and propose solutions to these issues which lead to a significant boost in attack performance for all SOTA MI. In particular, our contributions are two-fold: 1) We ana- lyze the optimization objective of SOTA MI algorithms, ar- gue that the objective is sub-optimal for achieving MI, and propose an improved optimization objective that boosts at- tack performance significantly. 2) We analyze “MI overfit- ting”, show that it would prevent reconstructed images from learning semantics of training data, and propose a novel “model augmentation” idea to overcome this issue. Our proposed solutions are simple and improve all SOTA MI at- tack accuracy significantly. E.g., in the standard CelebA benchmark, our solutions improve accuracy by 11.8% and achieve for the first time over 90% attack accuracy. Our findings demonstrate that there is a clear risk of leak- ing sensitive information from deep learning models. We urge serious consideration to be given to the privacy im- plications. Our code, demo, and models are available athttps://ngoc-nguyen-0.github.io/re- thinking_model_inversion_attacks/ .

1. Introduction Privacy of deep neural networks (DNNs) has attracted considerable attention recently [2, 3, 23, 31, 32]. Today, DNNs are being applied in many domains involving pri- vate and sensitive datasets, e.g., healthcare, and security. There is a growing concern of privacy attacks to gain knowl- edge of confidential datasets used in training DNNs. One *Equal Contribution†Corresponding Authorimportant category of privacy attacks is Model Inversion (MI) [5, 8, 11, 12, 16, 36, 37, 39, 40] (Fig. 1). Given ac- cess to a model, MI attacks aim to infer and reconstruct fea- tures of the private dataset used in the training of the model. For example, a malicious user may attack a face recognition system to reconstruct sensitive face images used in training. Similar to previous work [5,36,39], we will use face recog- nition models as the running example. Related Work. MI attacks were first introduced in [12], where simple linear regression is the target of attack. Re- cently, there is a fair amount of interest to extend MI to com- plex DNNs. Most of these attacks [5, 36, 39] focus on the whitebox setting and the attacker is assumed to have com- plete knowledge of the model subject to attack. As many platforms provide downloading of entire trained DNNs for users [5, 39], whitebox attacks are important. [39] proposes Generative Model Inversion (GMI) attack, where generic public information is leveraged to learn a distributional prior via generative adversarial networks (GANs) [13, 35], and this prior is used to guide reconstruction of private training samples. [5] proposes Knowledge-Enriched Dis- tributional Model Inversion (KEDMI), where an inversion- specific GAN is trained by leveraging knowledge provided by the target model. [36] proposes Variational Model Inver- sion (VMI), where a probabilistic interpretation of MI leads to a variational objective for the attack. KEDMI and VMI achieve SOTA attack performance (See Supplementary for further discussion of related work). In this paper , we revisit SOTA MI, study two issues pertaining to all SOTA MI and propose solutions to these is- sues that are complementary and applicable to all SOTA MI (Fig. 1). In particular, despite the range of approaches pro- posed in recent works, common and central to all these ap- proaches is an inversion step which formulates reconstruc- tion of training samples as an optimization. The optimiza- tion objective in the inversion step involves the identity loss , which is the same for all SOTA MI and is formulated as the negative log-likelihood for the reconstructed samples under the model being attacked. While ideas have been proposed to advance other aspects of MI, effective design of the iden- tity loss has not been studied . 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. 16384 An Improved

Qu_Modality-Agnostic_Debiasing_for_Single_Domain_Generalization_CVPR_2023

Abstract Deep neural networks (DNNs) usually fail to general- ize well to outside of distribution (OOD) data, especially in the extreme case of single domain generalization (single- DG) that transfers DNNs from single domain to multi-ple unseen domains. Existing single-DG techniques com-monly devise various data-augmentation algorithms, andremould the multi-source domain generalization methodol-ogy to learn domain-generalized (semantic) features. Nev-ertheless, these methods are typically modality-specific,thereby being only applicable to one single modality (e.g.,image). In contrast, we target a versatile Modality-AgnosticDebiasing (MAD) framework for single-DG, that enablesgeneralization for different modalities. Technically, MADintroduces a novel two-branch classifier: a biased-branch encourages the classifier to identify the domain-specific (su- perficial) features, and a general-branch captures domain-generalized features based on the knowledge from biased-branch. Our MAD is appealing in view that it is pluggableto most single-DG models. We validate the superiority of our MAD in a variety of single-DG scenarios with different modalities, including recognition on 1D texts, 2D images,3D point clouds, and semantic segmentation on 2D images.More remarkably, for recognition on 3D point clouds andsemantic segmentation on 2D images, MAD improves DSUby 2.82% and 1.5% in accuracy and mIOU.

1. Introduction Deep neural networks (DNNs) have achieved remarkable success in various tasks under the assumption that train-ing and testing domains are independent and sampled fromidentical or sufficiently similar distribution [ 2,48]. How- ever, this assumption often does not hold in most real-world scenarios. When deploying DNNs to unseen or out-of-distribution (OOD) testing domains, inevitable perfor-mance degeneration is commonly observed. The difficulty mainly originates from that the backbone of DNNs ex- *Corresponding authorSamples in Training Domain OOD Samples Data Augmentation Can not directly transfer ...Images Images Point CloudsPoint Clouds Figure 1. Most existing single-DG techniques devise various data augmentation algorithms to introduce various image texturesand styles, pursuing the learning of domain-generalized features.However, these approaches are modality-specific, and only appli-cable to single modality (e.g., image). Hence it is difficult to di-rectly employ such single-DG approach for 3D point clouds, sincethe domain shifts in 3D point clouds only reflect the geometric dif- ferences rather than texture and style differences. tracts more domain-specific (superficial) features together with domain-generalized (semantic) features. Therefore,the classifier is prone to paying much attention to thosedomain-specific features, and learning unintended decisionrule [ 53]. To mitigate this issue, several appealing so- lutions have been developed, including Domain Adapta- tion (DA) [18,32,36,40,41] and Domain Generalization (DG) [31,56,62,65]. Despite showing encouraging per- formances on OOD data, their real-world applications arestill limited due to the requirement to have the data from other domain (i.e., the unseen target domain or multiplesource domains with different distributions). In this work,we focus on an extreme case in domain generalization: sin- gle domain generalization (single-DG) , in which DNNs are trained with single source domain data and then required togeneralize well to multiple unseen target domains. Previous researches [ 19,55] demonstrate that the specific local textures and image styles tailored to each domain aretwo main causes, resulting in domain-specific features forimages. To alleviate this, recent works [ 30,37,58,63] 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. 24142 sign a variety of data-augmentation algorithms to introduce diversified textures and image styles. The DG methodolo-gies are then remolded with these data-augmentation algo-rithms to facilitate the learning of domain-generalized fea-tures. Nevertheless, such solution for single-DG is typicallymodality-specific and only applicable to the single modal-ity inputs of images. When coming a new modality (e.g.3D point clouds), it is difficult to directly apply these tech-niques to tackle single-DG problem. This is due to the factthat the domain shift in 3D point clouds is interpreted asthe differences of 3D structural information among multi- ple domains, instead of the texture and style differences in 2D images [ 10,39]. Figure 1conceptually illustrates the issue, which has been seldom explored in the literature. In this paper, we propose to address this limitation from the standpoint of directly strengthening the capac-ity of classifier to identify domain-specific features, and meanwhile emphasize the learning of domain-generalizedfeatures. Such way completely eliminates the need ofmodality-specific data augmentations, thereby leading to aversatile modality-agnostic paradigm for single-DG. Tech- nically, to materialize this idea, we design a novel Modality- Agnostic Debiasing (MAD) framework, that facilitates sin-gle domain generalization under a wide variety of modali-ties. In particular, MAD integrates the basic backbone forfeature extraction with a new two-branch classifier struc- ture. One branch is the biased-branch that identifies thosesuperficial and domain-specific features with a multi-head cooperated classifier. The other branch is the general-branch that learns to capture the domain-generalized rep-resentations on the basis of the knowledge derived from thebiased-branch. It is also appealing in view that our MAD can be seamlessly incorporated into most existing single- DG models with data-augmentation, thereby further boost-ing single domain generalization. We analyze and evaluate our MAD under a variety of single-DG scenarios with different modalities, ranging fromrecognition on 2D images, 3D point clouds, 1D texts, tosemantic segmentation on 2D images. Extensive experi- ments demonstrate the superior advantages of MAD whenbeing plugged into a series of existing single-DG techniqueswith data-augmentation (e.g., Mixstyle [ 65] and DSU [ 30]). More remarkably, for recognition on point cloud bench-mark, MAD significantly improves DSU in the accuracy from 33.63% to 36.45%. For semantic segmentation on im-age benchmark, MAD advances DSU with mIoU improve-ment from 42.3% to 43.8%.

Qiu_Graph_Representation_for_Order-Aware_Visual_Transformation_CVPR_2023

Abstract This paper proposes a new visual reasoning formula- tion that aims at discovering changes between image pairs andtheirtemporalorders. Recognizingscenedynamicsand theirchronologicalordersisafundamentalaspectofhuman cognition. The aforementioned abilities make it possible to follow step-by-step instructions, reason about and analyze events, recognize abnormal dynamics, and restore scenes to their previous states. However, it remains unclear how well current AI systems perform in these capabilities. Al- though a series of studies have focused on identifying and describing changes from image pairs, they mainly consider thosechangesthatoccursynchronously,thusneglectingpo- tential orders within those changes. To address the above issue, we first propose a visual transformation graph struc- ture for conveying order-aware changes. Then, we bench- marked previous methods on our newly generated dataset andidentifiedtheissuesofexistingmethodsforchangeorder recognition. Finally, we show a significant improvement in order-awarechangerecognitionbyintroducinganewmodel that explicitly associates different changes and then identi- fies changes and their orders in a graph representation.

1. Introduction The Only Constant in Life Is Change. - Heraclitus Humans conduct numerous reasoning processes beyond objectandmotionrecognition. Throughtheseprocesses,we cancaptureawiderangeofinformationwithjustaglimpse ofascenario. Toachievehuman-levelvisualunderstanding, variousstudieshaverecentlyfocusedondifferentaspectsof visualreasoning,suchascompositional[1–4],causal[5,6], abstract [7–9], abductive [10,11], and commonsense visual reasoning [12, 13]. Due to the ever-changing visual sur- rounding,perceivingandreasoningoverscenedynamicsare essential. However, most existing visual reasoning studies focus on scenes in fixed periods of time. Therefore, this study focuses on a new formulation of visual reasoning for identifying scene dynamics. Encoder VTGen Before-changeAfter-change Type: moveObj.: cyan, rubber, cubePos0.: purple, rubber, spherePos1.: groundType: addObj.: gray, rubber, cylinderPos0.: -Pos1.: cyan, rubber, cubeType: addObj.: gray, rubber, cubePos0.: -Pos1.: purple, rubber, sphereType: moveObj.: gray, metal, spherePos0.: groundPos1.: blue, rubber, sphere Visual transformation graphFigure 1. Overview of the proposed order-aware change recog- nition model VTGen (top). From an image pair observed before andaftermultiplesynchronousandasynchronouschanges,VTGen generates a visual transformation graph (bottom) where nodes in- dicate change contents (including type, object attributes, original positiondescribedbywhatisunderneathit,andnewposition,and directed edges indicate temporal orders of changes. Due to variations in the spatial positions of objects and the temporal order of human activities, changes within a pair of observations could occur simultaneously or asyn- chronously. Several recent studies have already discussed recognizinganddescribingsynchronouschangesfromapair ofimages vianatural languagetexts [14–16]while neglect- ing the potential orders between changes. However, iden- tifying temporal orders is an integral aspect of revealing how scene dynamics occur in time, making it possible to restore scenes to their previous states. Temporal orders are also critical in a variety of applications, such as room rear- rangement[17],assemblyoperation[18,19],andinstruction following [20,21]. Change order recognition presents new challenges as it requires reasoning over underlying tempo- 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. 22793 ralevents,anditalsocomplicatessinglechangerecognition due to entangled appearances and localization. However, despite its complexity, humans exhibit high performance in findingmultiplechangesanddeterminingtheirtemporalor- ders from a pair of images. For example, even human chil- dren under age five can perform assembly operations with toysingameslikeLEGObuildingblocks. Therefore,inthis work, we are particularly interested in how well the current AI methods perform in order-aware change recognition. Similartodiscoveringdynamicsandordersfromapairof scene observations, there is a group of works that discusses the step-by-step assembly of objects from their component parts [18,22–24]. Such part assembly studies tend to focus on recovering the sequence steps for rebuilding objects and are thus highly useful in robotic applications used for as- sembly operations or instruction following. However, part assembly operations focus on the reconstruction of objects fromtheirparts,andnotfindingthedifferencesbetweentwo discrete scene observations. Moreover, instead of directly recovering all steps from two single observations, existing partassemblymethodsrequireadditionalinformation,such as language instructions or demonstration videos, for their step generation processes. As shown in Figure 1, this study proposes a new task to identifyorder-awarechangesdirectlyfromapairofimages. Most existing studies generate a single sentence [14,15], paragraph[16],ortriplets[25]fordescribingchanges. How- ever, sentences are lengthy and less suitable for simultane- ouslyindicatingchangecontentsandtheirorders,andmake model analysis and evaluation opaque. Hence, we propose the use of an order-aware transformation graph (Figure 1 bottom). Changecontentsarerepresentedbynodesandtheir chronological orders by directed edges. To diagnose model performance,wegeneratedadataset,namedorder-awarevi- sualtransformation(OVT),consistingofasynchronousand synchronous changes between scene observations. We then conducted benchmark experiments using ex- isting methods and found they showed seriously degraded performance in terms of order-aware change recognition. Although neglected by existing methods, associations be- tween changes, and disentangled representations of change contents and orders are useful in identifying order-aware changes. Therefore, we propose a novel method called vi- sualtransformationgraphgenerator(VTGen)thatexplicitly associates different changes and generates a graph that de- scribes change contents and their orders in a disentangled manner. VTGen achieved state-of-the-art performance in theOVTdatasetandanexistingbenchmarkCLEVR-Multi- Change [16], and outperformed existing methods by large margins. However,wealsofoundasignificantperformance gap between the best-performing model and humans. We hopeourresearchandOVTdatasetcancontributetoachiev- ing human-level visual reasoning in scene dynamics.Our contributions are three-fold: i. We propose a novel taskandadatasetnamedOVTfororder-awarevisualtrans- formation. ii. We report on benchmark evaluations of ex- isting change recognition methods in order-aware change recognitionanddiscusstheirshortcomings. iii. Wepropose anovelmethodVTGenthatachievesstate-of-the-artperfor- mance in the OVT dataset and an existing change recogni- tion benchmark.

Metzger_Guided_Depth_Super-Resolution_by_Deep_Anisotropic_Diffusion_CVPR_2023

Abstract Performing super-resolution of a depth image using the guidance from an RGB image is a problem that con- cerns several fields, such as robotics, medical imaging, and remote sensing. While deep learning methods have achieved good results in this problem, recent work high- lighted the value of combining modern methods with more formal frameworks. In this work, we propose a novel ap- proach which combines guided anisotropic diffusion with a deep convolutional network and advances the state of the art for guided depth super-resolution. The edge transfer- ring/enhancing properties of the diffusion are boosted by the contextual reasoning capabilities of modern networks, and a strict adjustment step guarantees perfect adherence to the source image. We achieve unprecedented results in three commonly used benchmarks for guided depth super- resolution. The performance gain compared to other meth- ods is the largest at larger scales, such as ×32 scaling. Code1for the proposed method is available to promote re- producibility of our results.

1. Introduction It is a primordial need for visual data analysis to in- crease the resolution of images after they have been cap- tured. In many fields one is faced with images that, for technical reasons, have too low resolutions for the intended purposes, e.g., MRI scans in medical imaging [48], multi- spectral satellite images in Earth observation [22], thermal surveillance images [1] and depth images in robotics [9]. In some cases, an image of much higher resolution is available in a different imaging modality, which can act as a guide for super-resolving the low-resolution source image, by in- jecting the missing high-frequency content. For instance, in Earth observation, the guide is often a panchromatic im- age (hence the term ”pan-sharpening”), whereas in robotics a conventional RGB image is often attached to the same *Equal contribution. 1https://github.com/prs- eth/Diffusion- Super- Resolution Diffusion AdjustmentGuide Source InitializationDiffusion coef ficients Diffused result Adjusted result Diffusion-adjustment loop CNNFigure 1. We super-resolve a low-resolution depth image by find- ing the equilibrium state of a constrained anisotropic diffusion pro- cess. Learned diffusion coefficients favor smooth depth within ob- jects and suppress diffusion across discontinuities. They are de- rived from the guide with a neural feature extractor that is trained by back-propagating through the diffusion process. platform as a TOF camera or laser scanner. In this paper, we focus on super-resolving depth images guided by RGB images, but the proposed framework is generic and can be adapted to other sensor combinations, too. Research into guided super-resolution has a long his- tory [16, 29]. The proposed solutions range from classical, entirely hand-crafted schemes [10] to fully learning-based methods [15], while some recent works have combined the two schools of thought, with promising results [5, 32]. Many classical methods boil down to an image-specific op- timization problem that must be solved at inference time, which often makes them slow and memory-hungry. More- over, they are limited to low-level image properties of the guide, such as color and contrast, and lack the high-level image understanding and contextual reasoning of modern neural networks. On the positive side, by design, they can not overfit the peculiarities of a training set and tend to gen- eralize better. Recent work on guided super-resolution has focused on deep neural networks. Their superior ability to capture latent image structure has greatly advanced the state of the art over traditional, learning-free approaches. Still, these learning-based methods tend to struggle with sharp 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. 18237 discontinuities and often produce blurry edges in the super- resolved depth maps. Moreover, like many deep learning systems, they degrade – often substantially – when applied to images with even slightly different characteristics. Note also that standard feed-forward architectures cannot guar- antee a consistent solution: feeding the source and guiding images through an encoder-decoder structure to obtain the super-resolved target will, by itself, not ensure that down- sampling the target will reproduce the source. We propose a novel approach for guided depth super- resolution which combines the strengths of optimization- based and deep learning-based super-resolution. In short, our method is a combination of anisotropic diffusion (based on the discretized version of the heat equation) with deep feature learning (based on a convolutional backbone). The diffusion part resembles classical optimization approaches, solved via an iterative diffusion-adjustment loop. Every it- eration consists of (1) an anisotropic diffusion step [2, 4, 23, 30], with diffusion weights driven by the guide in such a way that diffusion (i.e., smoothing) is low across high- contrast boundaries and high within homogeneous regions; and (2) an adjustment step that rescales the depth values such that they exactly match the low-resolution source when downsampled. To harness the unmatched ability of deep learning to extract informative image features, the diffusion weights are not computed from raw brightness values but are set by passing the guide through a (fully) convolutional feature extractor. An overview of the method is depicted in Fig. 1. The technical core of our method is the insight that such a feature extractor can be trained end-to-end to optimally fulfill the requirements of the subsequent opti- mization, by back-propagating gradients through the iter- ations of the diffusion loop. Despite its apparent simplicity, this hybrid approach delivers excellent super-resolution re- sults. In our experiments, it consistently outperforms prior art on three different datasets, across a range of upsampling factors from ×4 to×32. In our experiments, we compare it to six recent learning methods as well as five different learning-free methods. For completeness, we also include a learning-free version of our diffusion-adjustment scheme and show that it outperforms all other learning-free meth- ods. Beyond the empirical performance, our method in- herits several attractive properties from its ingredients: the diffusion-based optimization scheme ensures strict adher- ence to the depth values of the source, crisp edges, and a degree of interpretability; whereas deep learning equips the very local diffusion weights with large-scale context infor- mation, and offers a tractable, constant memory footprint at inference time. In summary, our contributions are: 1. We develop a hybrid framework for guided super- resolution that combines deep feature learning and anisotropic diffusion in an integrated, end-to-end train- able pipeline;2. We provide an implementation of that scheme with constant memory demands, and with inference time that is constant for a given upsampling factor and scales linearly with the number of iterations; 3. We set a new state of the art for the Middlebury [38], NYUv2 [39] and DIML [20] datasets, for upsampling factors from 4 ×to 32×, and provide empirical evi- dence that our method indeed guarantees exact consis- tency with the source image.

Luo_Towards_Generalisable_Video_Moment_Retrieval_Visual-Dynamic_Injection_to_Image-Text_Pre-Training_CVPR_2023

Abstract The correlation between the vision and text is essen- tial for video moment retrieval (VMR), however, existing methods heavily rely on separate pre-training feature ex- tractors for visual and textual understanding. Without suf- ficient temporal boundary annotations, it is non-trivial to learn universal video-text alignments. In this work, we explore multi-modal correlations derived from large-scale image-text data to facilitate generalisable VMR. To address the limitations of image-text pre-training models on captur- ing the video changes, we propose a generic method, re- ferred to as Visual-Dynamic Injection (VDI), to empower the model’s understanding of video moments. Whilst ex- isting VMR methods are focusing on building temporal- aware video features, being aware of the text descriptions about the temporal changes is also critical but originally overlooked in pre-training by matching static images with sentences. Therefore, we extract visual context and spa- tial dynamic information from video frames and explicitly enforce their alignments with the phrases describing video changes ( e.g. verb). By doing so, the potentially relevant visual and motion patterns in videos are encoded in the cor- responding text embeddings (injected) so to enable more accurate video-text alignments. We conduct extensive ex- periments on two VMR benchmark datasets (Charades-STA and ActivityNet-Captions) and achieve state-of-the-art per- formances. Especially, VDI yields notable advantages when being tested on the out-of-distribution splits where the test- ing samples involve novel scenes and vocabulary.

1. Introduction Video moment retrieval (VMR) aims at locating a video moment by its temporal boundary in a long and untrimmed video according to a natural language sentence [3, 13]. It *Corresponding authors (a) Separated TrainingTextVision√√TextVision(b) Video Text Joint TrainingxTextMomentTextVideo√(d) Ours (c) Image Text Joint TrainingActionsActionsTextImage√xVisual EncoderTextual EncoderVisual-Dynamic Injection√ MomentTextA person opens the door xFigure 1. Contemporary methods lack moment-text correlations. Our method takes the advantage of image-text pre-trained models and learns moment-text correlations by visual-dynamic injection. is a critical task which has been extensively studied in a va- riety of real-world applications including human-computer interaction [5], and intelligent surveillance [9]. In practice, raw videos are usually unscripted and unstructured, while the words being chosen for describing the same video mo- ments can be varied from person to person [45, 63]. To be generalisable to different scenes, VMR is fundamentally challenging as it requires the comprehension of arbitrary complex visual and motion patterns in videos and an un- bounded vocabulary with their intricate relationships. For the fine-grained retrieval objective of VMR, the pre- cise segment-wise temporal boundary labels are intuitively harder to be collected than conventional image/video-level annotations. In this case, rather than training from scratch with a limited number of temporally labelled videos, ex- isting VMR solutions [3, 13, 14, 62] heavily rely on single- modal pre-training [8, 48] for visual and textual understand- ing (Fig. 1 (a)). By doing so, they focus on modelling the correlations between the pre-learned features of videos and sentences. Nonetheless, without sufficient training data, it is non-trivial to derive universal video-text alignments so 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. 23045 generalise to novel scenes and vocabulary. Separately, the recent successes achieved by joint vision- language pre-training in zero-shot learning [21, 42] demon- strate the potential of adapting the multi-modal correlations derived from large-scale visual-textual data to facilitate gen- eralisable VMR. Whilst it is intuitive to adopt the video- text pre-learned features [34, 38, 50] for moment retrieval (Fig. 1 (b)), it has been shown that the models pre-trained with coarse-grained video-level labels can not transfer well to localisation-based tasks like VMR due to their unaware- ness of fine-grained alignments between text and frames or clips [2]. Such a misalignment problem is less likely to exist in pre-training by image-text matching. However, image- based pre-training models [21, 42] are less sensitive to the changes in videos and the words describing such dynamics in text [17]. This is inherent in matching sentences and im- ages with static content but is significant in understanding video actions and activities (Fig. 1 (c)). It is suboptimal to directly apply image-text pre-learned features on VMR. In this work, we propose a generic method for exploit- ing large-scale image-text pre-training models to benefit generalisable VMR by the universal visual-textual corre- lations derived in pre-training, dubbed as Visual-Dynamic Injection (VDI). The key idea is to explore the visual con- text and spatial dynamic information from videos and in- ject that into text embeddings to explicitly emphasise the phrases describing video changes ( e.g. verb) in sentences (Fig. 1 (d)). Such visual and dynamic information in text is critical for locating video moments composed of arbitrary evolving events but unavailable or overlooked in image- text pre-training. Specifically, we consider it essential for VMR models to answer two questions: “what are the ob- jects” and “how do the objects change”. The visual context information indicates the content in the frames, e.g. back- grounds (scenes), appearances of objects, poses of subjects, etc. Meanwhile, the spatial dynamic is about the location changes of different salient entities in a video, which po- tentially implies the development of their interactions. VDI is a generic formulation, which can be integrated into any existing VMR model. The only refinement is to adapt the text encoder by visual-dynamic information injection dur- ing training. Hence, no additional computation costs are introduced in inference. Our contributions are three-folded: (1)To our best knowledge, this is the first attempt on injecting visual and dynamic information to image-text pre-training models to enable generalisable VMR. (2)We propose a novel method for VMR called Visual-Dynamic Injection (VDI). The VDI method is a generic formulation that can be integrated into existing VMR models and benefits them from the universal visual-textual alignments derived from large-scale image- text data. (3)The VDI achieves the state-of-the-art perfor- mances on two standard VMR benchmark datasets. Moreimportantly, it yields notable performance advantages when being tested on the out-of-distribution splits where the test- ing samples involve novel scenes and vocabulary. VDI’s superior generalisation ability demonstrates its potential for adapting image-text pre-training to video understanding tasks requiring fine-grained visual-textual comprehensions.

Qin_Learning_To_Exploit_the_Sequence-Specific_Prior_Knowledge_for_Image_Processing_CVPR_2023

Abstract The hardware image signal processing (ISP) pipeline is the intermediate layer between the imaging sensor and the downstream application, processing the sensor signal into an RGB image. The ISP is less programmable and con- sists of a series of processing modules. Each processing module handles a subtask and contains a set of tunable hy- perparameters. A large number of hyperparameters form a complex mapping with the ISP output. The industry typi- cally relies on manual and time-consuming hyperparameter tuning by image experts, biased towards human perception. Recently, several automatic ISP hyperparameter optimiza- tion methods using downstream evaluation metrics come into sight. However, existing methods for ISP tuning treat the high-dimensional parameter space as a global space for optimization and prediction all at once without inducing the structure knowledge of ISP . To this end, we propose a se- quential ISP hyperparameter prediction framework that uti- lizes the sequential relationship within ISP modules and the similarity among parameters to guide the model sequence process. We validate the proposed method on object detec- tion, image segmentation, and image quality tasks.

1. Introduction Hardware ISPs are low-level image processing pipelines that convert RAW images into high-quality RGB images. Typically, ISPs include a series of processing modules [5], each of which handles a subtask such as denoising, white balance, demosaicing, or sharpening. Compared to soft- ware image processing pipelines, hardware ISPs are faster, more power-efficient, and widely used in real-time prod- ucts [7, 34], including cameras [28], smartphones, surveil-lance [21], IoT and driven-assistance systems. ISPs are highly modular and less programmable but with a set of tun- able hyperparameters. The industry always relies on man- ual and costly hyperparameter tuning by image experts [1] to adapt the ISP to different application scenarios. The ISP is always designed as a sequential pipeline [5] and the configurable hyperparameters of various modules from any ISP aggregate to be a complex parameter space (with tens to hundreds of parameters), making the manual tuning process time-consuming. It is also difficult to sub- jectively find optimal hyperparameters settings for various downstream tasks (such as object detection and image seg- mentation [29, 30]). Recently, several automatic ISP hyper- parameter optimization methods [17,31] using downstream evaluation metrics come into sight. These methods tuning hyperparameters for downstream tasks include derivative- free [18] or gradient methods [12, 25, 27] based on differ- entiable approximation. There are also methods to demon- strate the potential of predicting specific hyperparameters for each image or scene [22]. However, existing meth- ods treat the high-dimensional parameter space as a global black-box space for optimization and prediction all at once, while ignoring the inherent sequence of the ISP modules and the critical intra-correlation among hyperparameters. Inspired by the operating principles and structure knowl- edge of ISPs, we first propose a sequential ISP hyperpa- rameter prediction framework (as shown in Fig. 1) that con- tains Sequential CNN model and Global Similarity Group- ing. The sequential CNN model runs recurrently by predict- ing a group of parameters from several ISP modules, not all parameters, at each step. Meanwhile, the predicted param- eters, along with the network’s hidden layer and the input data, are in turn encoded as prior knowledge for predicting the following grouping of parameters. The Global Similar- ity Grouping module divides ISP parameters into multiple 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. 22314 ISP Hyperparameter Space ISP Hyperparameter Space Sequential Process Block1 Block2 Block3 Block4 Block5 Image Processing Pipelines (ISP)RAW InputRGB OutputBlock1 Block2 Block3 Block4 Block5 Image Processing Pipelines (ISP)RAW InputRGB OutputExisting Methods (A) Proposed Method (B) SimilarityGroupingOptimize All at Once ConcatenationFigure 1. (A) Previous methods treat the hyperparameter space as a black box optimization problem and estimate all parameters at once without considering the prior knowledge of the ISP. (B) Our proposed method first decouples ISP structural knowledge and treats ISP tuning as a sequential prediction problem. It is effective to introduce sequence information and similarity relations in the high-dimensional hyperparameter space. disjoint groupings. The sequence order between groupings is explored heuristically using prior knowledge of the order of ISP modules. Given the flexibility, groupings are deter- mined based on similarity among parameters, not limited to the same module parameters. The correlation of parameter activation maps learned through the model is used as the basis for parameter groupings. Our contributions can be summarized as the following: • We propose a new sequential CNN structure to exploit the sequence processing knowledge within ISP. The potential sequential information among parameters is used to guide the processing of the model. • We exploit the correlation among parameters by the proposed similarity grouping module. The flexible parameter groupings allow the exploration of cross- module relationships among parameters. • We validate the effectiveness of our method in a variety of downstream tasks, including object detection, image segmentation, and image quality. In these applications,our method outperforms existing methods.

Li_SViTT_Temporal_Learning_of_Sparse_Video-Text_Transformers_CVPR_2023

Abstract Do video-text transformers learn to model temporal re- lationships across frames? Despite their immense capacity and the abundance of multimodal training data, recent work has revealed the strong tendency of video-text models to- wards frame-based spatial representations, while temporal reasoning remains largely unsolved. In this work, we iden- tify several key challenges in temporal learning of video- text transformers: the spatiotemporal trade-off from limited network size; the curse of dimensionality for multi-frame modeling; and the diminishing returns of semantic informa- tion by extending clip length. Guided by these findings, we propose SViTT , a sparse video-text architecture that per- forms multi-frame reasoning with significantly lower cost than na ¨ıve transformers with dense attention. Analogous to graph-based networks, SViTT employs two forms of sparsity: edge sparsity that limits the query-key commu- nications between tokens in self-attention, and node spar- sity that discards uninformative visual tokens. Trained with a curriculum which increases model sparsity with the clip length, SViTT outperforms dense transformer baselines on multiple video-text retrieval and question answering bench- marks, with a fraction of computational cost. Project page: http://svcl.ucsd.edu/projects/svitt .

1. Introduction With the rapid development of deep neural networks for computer vision and natural language processing, there has been growing interest in learning correspondences across the visual and text modalities. A variety of vision-language pretraining frameworks have been proposed [12, 22, 29, 34] for learning high-quality cross-modal representations with weak supervision. Recently, progress on visual transform- ers (ViT) [5,16,32] has enabled seamless integration of both modalities into a unified attention model, leading to image- text transformer architectures that achieve state-the-art per- formance on vision-language benchmarks [1, 27, 44]. Progress has also occurred in video -language pretraining by leveraging image-text models for improved frame-based *Work done during an internship at Intel Labs. Query Key / V alue Edge sparsity Node sparsity One person takes a pillow and some homework from an old chair , then smiles and laughs.Figure 1. We propose SViTT , asparse video-text transformer for efficient modeling of temporal relationships across video frames. Top: Semantic information for video-text reasoning is highly lo- calized in the spatiotemporal volume, making dense modeling in- efficient and prone to contextual noises. Bottom :SViTT pur- sues edge sparsity by limiting query-key pairs in self-attention, and node sparsity by pruning redundant tokens from visual sequence. reasoning [4, 9, 18]. Spatial modeling has the advantage of efficient (linear) scaling to long duration videos. Per- haps due to this, single-frame models have proven surpris- ingly effective at video-text tasks, matching or exceeding prior arts with complex temporal components [9,24]. How- ever, spatial modeling creates a bias towards static appear- ance and overlooks the importance of temporal reasoning in videos. This suggests the question: Are temporal dynamics not worth modeling in the video-language domain? Upon a closer investigation, we identify a few key chal- lenges to incorporating multi-frame reasoning in video- language models. First, limited model size implies a trade- off between spatial and temporal learning (a classic example being 2D/3D convolutions in video CNNs [46]). For any given dataset, optimal performance requires a careful bal- 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. 18919 ance between the two. Second, long-term video models typ- ically have larger model sizes and are more prone to over- fitting. Hence, for longer term video models, it becomes more important to carefully allocate parameters and control model growth. Finally, even if extending the clip length im- proves the results, it is subject to diminishing returns since the amount of information provided by a video clip does not grow linearly with its sampling rate. If the model size is not controlled, the computational increase may not jus- tify the gains in accuracy. This is critical for transformer- based architectures, since self-attention mechanisms have a quadratic memory and time cost with respect to input length. In summary, model complexity should be adjusted adaptively, depending on the input videos, to achieve the best trade-off between spatial representation, temporal rep- resentation, overfitting potential, and complexity. Since ex- isting video-text models lack this ability, they either attain a suboptimal balance between spatial and temporal modeling, or do not learn meaningful temporal representations at all. Motivated by these findings, we argue that video-text models should learn to allocate modeling resources to the video data. We hypothesize that, rather than uniformly ex- tending the model to longer clips, the allocation of these re- sources to the relevant spatiotemporal locations of the video is crucial for efficient learning from long clips. For trans- former models, this allocation is naturally performed by pruning redundant attention connections. We then propose to accomplish these goals by exploring transformer spar- sification techniques. This motivates the introduction of a Sparse Video-Text Transformer (SViTT ) inspired by graph models. As illustrated in Fig. 1, SViTT treats video to- kens as graph vertices, and self-attention patterns as edges that connect them. We design SViTT to pursue sparsity for both: edge sparsity aims at reducing query-key pairs in attention module while maintaining its global reasoning ca- pability; node sparsity reduces to identifying informative to- kens (e.g., corresponding to moving objects or person in the foreground) and pruning background feature embeddings. To address the diminishing returns for longer input clips, we propose to train SViTT with temporal sparse expansion , a curriculum learning strategy that increases clip length and model sparsity, in sync, at each training stage. SViTT is evaluated on diverse video-text benchmarks from video retrieval to question answering, comparing to prior arts and our own dense modeling baselines. First, we perform a series of ablation studies to understand the bene- fit of sparse modeling in transformers. Interestingly, we find that both nodes (tokens) and edges (attention) can be pruned drastically at inference, with a small impact on test perfor- mance. In fact, token selection using cross-modal attention improves retrieval results by 1% without re-training. We next perform full pre-training with the sparse mod- els and evaluate their downstream performance. We observethatSViTT scales well to longer input clips where the accu- racy of dense transformers drops due to optimization diffi- culties. On all video-text benchmarks, SViTT reports com- parable or better performance than their dense counterparts with lower computational cost, outperforming prior arts in- cluding those trained with additional image-text corpora. The key contributions of this work are: 1) a video-text architecture SViTT that unifies edge and node sparsity; 2) a sparse expansion curriculum for training SViTT on long video clips; and 3) empirical results that demonstrate its temporal modeling efficacy on video-language tasks.

Luo_GradMA_A_Gradient-Memory-Based_Accelerated_Federated_Learning_With_Alleviated_Catastrophic_Forgetting_CVPR_2023

Abstract Federated Learning (FL) has emerged as a de facto ma- chine learning area and received rapid increasing research interests from the community. However, catastrophic forget- ting caused by data heterogeneity and partial participation poses distinctive challenges for FL, which are detrimental to the performance. To tackle the problems, we propose a new FL approach (namely GradMA), which takes inspira- tion from continual learning to simultaneously correct the server-side and worker-side update directions as well as take full advantage of server’s rich computing and mem- ory resources. Furthermore, we elaborate a memory reduc- tion strategy to enable GradMA to accommodate FL with a large scale of workers. We then analyze convergence of GradMA theoretically under the smooth non-convex setting and show that its convergence rate achieves a linear speed up w.r.t the increasing number of sampled active workers. At last, our extensive experiments on various image classi- fication tasks show that GradMA achieves significant per- formance gains in accuracy and communication efficiency compared to SOTA baselines. We provide our code here: https://github.com/lkyddd/GradMA.

1. Introduction Federated Learning (FL) [18, 26] is a privacy-preserving distributed machine learning scheme in which workers jointly participate in the collaborative training of a central- ized model by sharing model information (parameters or updates) rather than their private datasets. In recent years, FL has shown its potential to facilitate real-world appli- cations, which falls broadly into two categories [10]: the cross-silo FL and the cross-device FL. The cross-silo FL corresponds to a relatively small number of reliable work- ers, usually organizations, such as healthcare facilities [9] and financial institutions [41], etc. In contrast, for the cross- *Corresponding authordevice FL, the number of workers can be very huge and unreliable, such as mobile devices [26], IoT [27] and au- tonomous driving cars [22], among others. In this paper, we focus on cross-device FL. The privacy-preserving and communication-efficient properties of the cross-device FL make it promising, but it also confronts practical challenges arising from data hetero- geneity (i.e., non-iid data distribution across workers) and partial participation [5,12,20,39]. Specifically, the datasets held by real-world workers are generated locally accord- ing to their individual circumstances, resulting in the dis- tribution of data on different workers being not identical. Moreover, owing to the flexibility of worker participation in many scenarios (e.g., IoT and mobile devices), workers can join or leave the FL system at will, thus making the set of active workers random and time-varying across commu- nication rounds. Note that we consider a worker participates or is active at round t(i.e., the index of the communication round) if it is able to complete the computation task and send back model information at the end of round t. The above-mentioned challenges mainly bring catas- trophic forgetting (CF) [25, 30, 37] to FL. In a typical FL process, represented by FedAvg [26], a server updates the centralized model by iteratively aggregating the model in- formation from workers that generally is trained over sev- eral steps locally before being sent to the server. On the one hand, due to data heterogeneity, the model is updated on private data in local training, which is prone to overfit the current knowledge and forget the previous experience, thus leading to CF [8]. In other words, the updates of the lo- cal models are prone to drift and diverge increasingly from the update of the centralized model [12]. This can seriously deteriorate the performance of the centralized model. To ameliorate this issue, a variety of existing efforts regular- ize the objectives of the local models to align the central- ized optimization objective [1, 12, 13, 17, 19]. On the other hand, the server can only aggregate model information from active workers per communication round caused by partial participation. In this case, many existing works directly dis- 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. 3708 card [1, 11, 12, 18, 26, 39] or implicitly utilize [7, 28], by means of momentum, the information provided by work- ers who have participated in the training but dropped out in the current communication round (i.e., stragglers). This results the centralized model, which tends to forget the ex- perience of the stragglers, thus inducing CF. In doing so, the convergence of popular FL approaches (e.g., FedAvg) can be seriously slowed down by stragglers. Moreover, all above approaches solely aggregate the collected informa- tion by averaging in the server, ignoring the server’s rich computing and memory resources that could be potentially harnessed to boost the performance of FL [45]. In this paper, to alleviate CF caused by data heterogene- ity and stragglers, we bring forward a new FL approach, dubbed as GradMA ( Grad ient-Memory-based Accelerated Federated Learning), which takes inspiration from contin- ual learning (CL) [4,14,24,29,44] to simultaneously correct the server-side and worker-side update directions and fully utilize the rich computing and memory resources of the server. Concretely, motivated by the success of GEM [24] and OGD [4], two memory-based CL methods, we invoke quadratic programming (QP) and memorize updates to cor- rect the update directions. On the worker side, GradMA harnesses the gradients of the local model in the previous step and the centralized model, and the parameters differ- ence between the local model in the current step and the centralized model as constraints of QP to adaptively correct the gradient of the local model. Furthermore, we maintain a memory state to memorize accumulated update of each worker on the server side. GradMA then explicitly takes the memory state to constrain QP to augment the momen- tum (i.e., the update direction) of the centralized model. Here, we need the server to allocate memory space to store memory state. However, it may be not feasible in FL scenar- ios with a large size of workers, which can increase the stor- age cost and the burden of computing QP largely. There- fore, we carefully craft a memory reduction strategy to alle- viate the said limitations. In addition, we theoretically ana- lyze the convergence of GradMA in the smooth non-convex setting. To sum up, we highlight our contributions as follows: • We formulate a novel FL approach GradMA, which aims to simultaneously correct the server-side and worker-side update directions and fully harness the server’s rich computing and memory resources. Mean- while, we tailor a memory reduction strategy for GradMA to reduce the scale of QP and memory cost. • For completeness, we analyze the convergence of GradMA theoretically in the smooth non-convex set- ting. As a result, the convergence result of GradMA achieves the linear speed up as the number of selected active workers increases.• We conduct extensive experiments on four com- monly used image classification datasets (i.e., MNIST, CIFAR-10, CIFAR-100 and Tiny-Imagenet) to show that GradMA is highly competitive compared with other state-of-the-art baselines. Meanwhile, ablation studies demonstrate efficacy and indispensability for core modules and key parameters.

Pu_Dynamic_Conceptional_Contrastive_Learning_for_Generalized_Category_Discovery_CVPR_2023

Abstract Generalized category discovery (GCD) is a recently pro- posed open-world problem, which aims to automatically cluster partially labeled data. The main challenge is that the unlabeled data contain instances that are not only from known categories of the labeled data but also from novel categories. This leads traditional novel category discov- ery (NCD) methods to be incapacitated for GCD, due to their assumption of unlabeled data are only from novel categories. One effective way for GCD is applying self- supervised learning to learn discriminate representation for unlabeled data. However, this manner largely ignores un- derlying relationships between instances of the same con- cepts (e.g., class, super-class, and sub-class), which re- sults in inferior representation learning. In this paper, we propose a Dynamic Conceptional Contrastive Learn- ing (DCCL) framework, which can effectively improve clus- tering accuracy by alternately estimating underlying vi- sual conceptions and learning conceptional representation. In addition, we design a dynamic conception generation and update mechanism, which is able to ensure consis- tent conception learning and thus further facilitate the opti- mization of DCCL. Extensive experiments show that DCCL achieves new state-of-the-art performances on six generic and fine-grained visual recognition datasets, especially on fine-grained ones. For example, our method significantly surpasses the best competitor by 16.2% on the new classes for the CUB-200 dataset. Code is available at https: //github.com/TPCD/DCCL

1. Introduction Learning recognition models ( e.g., image classification) from labeled data has been widely studied in the field of machine learning and deep learning [13, 18, 31]. In spite of their tremendous success, supervised learning techniques *Corresponding Author. AnimalsTransportationSuper-classRepresentationPulling FeaturePushingRepresentationSub-ClassDistributionRabbitPorcupine ClassDistribution Whale Bus UnlabbelledFeatureBicycle Train Sub-classRepresentationLabbelledFeature YellowTrainWhiteTrainFigure 1. Diagram of the proposed Dynamic Conceptional Con- trastive Learning (DCCL). Samples from the conceptions should be close to each other. For example, samples from the same classes (bus) at the class level, samples belonging to the transportation (bus and bicycle) at the super-class level, and samples from trains with different colors at the sub-class level. Our DCCL potentially learns the underlying conceptions in unlabeled data and produces more discriminative representations. rely heavily on huge annotated data, which is not suit- able for open-world applications. Thus, the researchers recently have paid much effort on learning with label- imperfection data, such as semi-supervised learning [23, 33], self-supervised learning [12, 42], weakly-supervised learning [41,45], few-shot learning [32,38], open-set recog- nition [30] and learning with noisy labels [40], etc. Recently, inspired by the fact that Humans can easily and automatically learn new knowledge with the guidance of previously learned knowledge, novel category discov- ery (NCD) [9, 11, 28, 44, 47] is introduced to automatically cluster unlabeled data of unseen categories with the help of knowledge from seen categories. However, the implemen- tation of NCD is under a strong assumption that all the un- labeled instances belong to unseen categories, which is not practical in real-world applications. To address this limi- tation, Vaze et al. [35] extend NCD to the generalized cat- egory discovery (GCD) [35], where unlabeled images are 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. 7579 from both novel and labeled categories. GCD is a challenging open-world problem in that we need to 1) jointly distinguish the known and unknown classes and 2) discover the novel clusters without any anno- tations. To solve this problem, Vaze et al. [35] leverage the contrastive learning technique to learn discriminative repre- sentation for unlabeled data and use k-means [21] to obtain final clustering results. In this method, the labeled data are fully exploited by supervised contrastive learning. How- ever, self-supervised learning is applied to the unlabeled data, which enforces samples to be close to their augmen- tation counterparts while far away from others. As a con- sequence, the underlying relationships between samples of the same conceptions are largely overlooked and thus will lead to degraded representation learning. Intuitively, sam- ples that belong to the same conceptions should be similar to each other in the feature space. The conceptions can be regarded as: classes, super-classes, sub-classes, etc. For ex- ample, as shown in Fig. 1, samples of the same class should be similar to each other, e.g., samples of the bus, samples of the bicycle. In addition, in the super-classes view, classes of the transportation, e.g., Bus and Bicycle, should belong to the same concept. Hence, the samples of transportation should be closer than that of other concepts ( e.g., animals). Similarly, samples belong to the same sub-classes ( e.g., red train) should be closer to that of other sub-classes ( e.g., white train). Hence, embracing such conceptions and their relationships can greatly benefit the representation learning for unlabeled data, especially for unseen classes. Motivated by this, we propose a Dynamic Conceptional Contrastive Learning (DCCL) framework for GCD to ef- fectively leverage the underlying relationships between un- labeled data for representation learning. Specifically, our DCCL includes two steps: Dynamic Conception Genera- tion (DCG) and Dual-level Contrastive Learning (DCL). In DCG, we dynamically generate conceptions based on the hyper-parameter-free clustering method equipped with the proposed semi-supervised conceptional consolidation. In DCL, we propose to optimize the model with conception- level and instance-level contrastive learning objectives, where we maintain a dynamic memory to ensure compar- ing with the up-to-date conceptions. The DCG and DCL are alternately performed until the model converges. We summarize the contributions of this work as follows: • We propose a novel dynamic conceptional contrastive learning (DCCL) framework to effectively leverage the underlying relationships between unlabeled samples for learning discriminative representation for GCD. • We introduce a novel dynamic conception generation and update mechanism to ensure consistent conception learning, which encourages the model to produce more discriminative representation.• Our DCCL approach consistently achieves superior performance over state-of-the-art GCD algorithms on both generic and fine-grained tasks.

Qiu_PSVT_End-to-End_Multi-Person_3D_Pose_and_Shape_Estimation_With_Progressive_CVPR_2023

Abstract Existing methods of multi-person video 3D human Pose and Shape Estimation (PSE) typically adopt a two-stage strategy, which first detects human instances in each frame and then performs single-person PSE with temporal model. However, the global spatio-temporal context among spa- tial instances can not be captured. In this paper, we pro- pose a new end-to-end multi-person 3D Pose and Shape estimation framework with progressive VideoTransformer, termed PSVT. In PSVT, a spatio-temporal encoder (STE) captures the global feature dependencies among spatial ob- jects. Then, spatio-temporal pose decoder (STPD) and shape decoder (STSD) capture the global dependencies be- tween pose queries and feature tokens, shape queries and feature tokens, respectively. To handle the variances of ob- jects as time proceeds, a novel scheme of progressive de- coding is used to update pose and shape queries at each frame. Besides, we propose a novel pose-guided attention (PGA) for shape decoder to better predict shape parame- ters. The two components strengthen the decoder of PSVT to improve performance. Extensive experiments on the four datasets show that PSVT achieves stage-of-the-art results.

1. Introduction Multi-person 3D human Pose and Shape Estimation (PSE) from monocular video requires localizing the 3D joint coordinates of all persons and reconstructing their human meshes (e.g. SMPL [29] model). As an essen- tial task in computer vision, it has many applications in- cluding human-robot interaction detection [22], virtual re- ality [33], and human behavior understanding [8], etc. Al- though remarkable progress has been achieved in PSE from videos [4, 41, 50, 51] or images [5, 44, 45], capturing multi- person spatio-temporal relations of pose and shape simulta- neously is still challenging since the difficulty in modeling long-range global interactions. …1T …1T …………HumanDetectionSPSESPSE STE…STD(a)Multi-stageframeworkofmulti-personPSEinvideo (b)End-to-endframeworkofPSVTMerging…PersonNPerson1 …T1 …T1Figure 1. Comparison of multi-stage and end-to-end framework. (a) Existing video-based methods [4, 16, 49, 50] perform single- person pose and shape estimation (SPSE) on the cropped areas by temporal modeling, such as Gated Recurrent Units (GRUs). (b) PSVT achieves end-to-end multi-person pose and shape estimation in video with spatial-temporal encoder (STE) and decoder (STD). To tackle this challenge, as shown in Figure 1 (a), exist- ing methods [4,16,50,51] employ a detection-based strategy of firstly detecting each human instance, then cropping the instance area in each frame and feeding it into the tempo- ral model, such as the recurrent neural network [4, 6, 16]. However, this framework can not capture the spatial re- lationship among human instances in an image and has the limitation of extracting long-range global context. Be- sides, the computational cost is expensive since it is pro- portional to the number of instances in image and it needs extra tracker [50] to identify each instance. Other temporal smoothing methods [15, 47] adopt a post-processing mod- ule to align the shape estimated by image-based PSE 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. 21254 proaches [5, 15, 23, 44, 45]. However, they can not capture temporal information directly from visual image features and lack the ability of long-range global interactions. These multi-stage methods split space and time dimensions and can not be end-to-end optimized. To strengthen the long-range modeling ability, recently developed Transformer models [7,46] have been introduced in PSE. The Transformer-based mesh reconstruction ap- proaches [24, 25, 36, 54] take each human joint as a to- ken and capture the relationship of human joints by atten- tion mechanism. However, the global context among dif- ferent persons in spatio-temporal dimensions has not been explored. Other Transformer-based human pose estimation approaches [27, 57] explore the spatio-temporal context of human joints for single-person, but not on the multi-person mesh. Besides, these methods focus on capturing the rela- tions among human joints, while ignoring the relations be- tween human poses and shapes. To tackle the above problems, we propose an end-to- end multi-person 3D Pose and Shape estimation framework with Video Transformer, termed PSVT, to capture long- range spatio-temporal global interactions in the video. As shown in Figure 1 (b), PSVT formulates the human in- stance localization and fine-grained pose and mesh estima- tion as a set prediction problem as [3, 42]. First, PSVT extracts a set of spatio-temporal tokens from the deep vi- sual features and applies a spatio-temporal encoder (STE) on these visual tokens to learn the relations of feature to- kens. Second, given a set of pose queries, a progressive spatio-temporal pose decoder (STPD) learns to capture the relations of human joints in both spatial and temporal di- mensions. Third, with the guidance of pose tokens from STPD, a progressive spatio-temporal shape decoder (STSD) learns to reason the relations of human mesh and pose in both spatial and temporal dimensions and further estimates the sequence 3D human mesh based on the spatio-temporal global context. Compared with previous shape estimation works [4,5,44,45,50,51], PSVT achieves end-to-end multi- person 3D pose and shape estimation in video. In PSVT, different from previous methods, we propose a novel progressive decoding mechanism (PDM) for se- quence decoding and pose-guided attention (PGA) for de- coder. PDM takes the output tokens from the last frame as the initialized queries for next frame, which enables bet- ter sequence decoding for STPD and STSD. PGA aligns the pose tokens and shape queries and further computes the cross-attention with feature tokens from encoder. With the guidance of pose tokens, shape estimation can be more ac- curate. Our contributions can be summarized as follows: We propose a novel video Transformer framework, termed PSVT, which is the first end-to-end multi- person 3D human pose and shape estimation frame- work with video Transformer.We propose a novel progressive decoding mechanism (PDM) for the decoder of video Transformer, which updates the queries at each frame in the attention block to improve the pose and shape decoding. We propose a novel pose-guided attention (PGA), which can capture the spatio-temporal relations among pose tokens, shape tokens, and feature tokens to im- prove the performance of shape estimation. Extensive experiments on four benchmarks show that PSVT achieves new state-of-the-art results.

Metaxas_DivClust_Controlling_Diversity_in_Deep_Clustering_CVPR_2023

Abstract Clustering has been a major research topic in the field of machine learning, one to which Deep Learning has re- cently been applied with significant success. However, an aspect of clustering that is not addressed by existing deep clustering methods, is that of efficiently producing multi- ple, diverse partitionings for a given dataset. This is par- ticularly important, as a diverse set of base clusterings are necessary for consensus clustering, which has been found to produce better and more robust results than relying on a single clustering. To address this gap, we propose Div- Clust, a diversity controlling loss that can be incorporated into existing deep clustering frameworks to produce mul- tiple clusterings with the desired degree of diversity. We conduct experiments with multiple datasets and deep clus- tering frameworks and show that: a) our method effectively controls diversity across frameworks and datasets with very small additional computational cost, b) the sets of clus- terings learned by DivClust include solutions that signifi- cantly outperform single-clustering baselines, and c) using an off-the-shelf consensus clustering algorithm, DivClust produces consensus clustering solutions that consistently outperform single-clustering baselines, effectively improv- ing the performance of the base deep clustering frame- work. Code is available at https://github.com/ ManiadisG/DivClust .

1. Introduction The exponentially increasing volume of visual data, along with advances in computing power and the develop- ment of powerful Deep Neural Network architectures, have revived the interest in unsupervised learning with visual data. Deep clustering in particular has been an area where significant progress has been made in the recent years. Ex- isting works focus on producing a single clustering, which is evaluated in terms of how well that clustering matches *Corresponding authorthe ground truth labels of the dataset in question. However, consensus, or ensemble, clustering remains under-studied in the context of deep clustering, despite the fact that it has been found to consistently improve performance over single clustering outcomes [3, 17, 45, 73]. Consensus clustering consists of two stages, specifically generating a set of base clusterings, and then applying a consensus algorithm to aggregate them. Identifying what properties ensembles should have in order to produce better outcomes in each setting has been an open problem [18]. However, research has found that inter-clustering diver- sity within the ensemble is an important, desirable fac- tor [14,20,24,34,51], along with individual clustering qual- ity, and that diversity should be moderated [15,22,51]. Fur- thermore, several works suggest that controlling diversity in ensembles is important toward studying its impact and determining its optimal level in each setting [22, 51]. The typical way to produce diverse clusterings is to pro- mote diversity by clustering the data multiple times with different initializations/hyperparameters or subsets of the data [3, 17]. This approach, however, does not guaran- tee or control the degree of diversity, and is computation- ally costly, particularly in the context of deep clustering, where it would require the training of multiple models. Some methods have been proposed that find diverse clus- terings by including diversity-related objectives to the clus- tering process, but those methods have only been applied to clustering precomputed features and cannot be trivially incorporated into Deep Learning frameworks. Other meth- ods tackle diverse clustering by creating and clustering di- verse feature subspaces, including some that apply this ap- proach in the context of deep clustering [48, 61]. Those methods, however, do not control inter-clustering diversity. Rather, they influence it indirectly through the properties of the subspaces they create. Furthermore, typically, exist- ing methods have been focusing on producing orthogonal clusterings or identifying clusterings based on independent attributes of relatively simple visual data (e.g. color/shape). Consequently, they are oriented toward maximizing inter- clustering diversity, which is not appropriate for consensus 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. 3418 Figure 1. Overview of DivClust. Assuming clusterings AandB, the proposed diversity loss Ldivcalculates their similarity matrix SAB and restricts the similarity between cluster pairs to be lower than a similarity upper bound d. In the figure, this is represented by the model adjusting the cluster boundaries to produce more diverse clusterings. Best seen in color. clustering [15, 22, 51]. To tackle this gap, namely generating multiple cluster- ings with deep clustering frameworks efficiently and with the desired degree of diversity, we propose DivClust. Our method can be straightforwardly incorporated into exist- ing deep clustering frameworks to learn multiple cluster- ings whose diversity is explicitly controlled . Specifically, the proposed method uses a single backbone for feature ex- traction, followed by multiple projection heads, each pro- ducing cluster assignments for a corresponding clustering. Given a user defined diversity target, in this work expressed in terms of the average NMI between clusterings, DivClust restricts inter-clustering similarity to be below an appropri- ate, dynamically estimated threshold. This is achieved with a novel loss component, which estimates inter-clustering similarity based on soft cluster assignments produced by the model, and penalizes values exceeding the threshold. Im- portantly, DivClust introduces minimal computational cost and requires no hyperparameter tuning with respect to the base deep clustering framework, which makes its use sim- ple and computationally efficient. Experiments on four datasets (CIFAR10, CIFAR100, Imagenet-10, Imagenet-Dogs) with three recent deep clus- tering methods (IIC [37], PICA [32], CC [44]) show that DivClust can effectively control inter-clustering diversity without reducing the quality of the clusterings. Further- more, we demonstrate that, with the use of an off-the-shelf consensus clustering algorithm, the diverse base clusterings learned by DivClust produce consensus clustering solutions that outperform the base frameworks, effectively improving them with minimal computational cost. Notably, despite the sensitivity of consensus clustering to the properties of the ensemble, our method is robust across various diversity lev- els, outperforming baselines in most settings, often by large margins. Our work then provides a straightforward way for improving the performance of deep clustering frameworks, as well as a new tool for studying the impact of diversity in deep clustering ensembles [51]. In summary, DivClust: a) can be incorporated in ex-isting deep clustering frameworks in a plug-and-play way with very small computational cost, b) can explicitly and effectively control inter-clustering diversity to satisfy user- defined targets, and c) learns clusterings that can improve the performance of deep clustering frameworks via consen- sus clustering.

Qin_Reliable_and_Interpretable_Personalized_Federated_Learning_CVPR_2023

Abstract Federated learning can coordinate multiple users to par- ticipate in data training while ensuring data privacy. The collaboration of multiple agents allows for a natural con- nection between federated learning and collective intelli- gence. When there are large differences in data distri- bution among clients, it is crucial for federated learning to design a reliable client selection strategy and an inter- pretable client communication framework to better utilize group knowledge. Herein, a reliable personalized federated learning approach, termed RIPFL, is proposed and fully in- terpreted from the perspective of social learning. RIPFL reliably selects and divides the clients involved in training such that each client can use different amounts of social information and more effectively communicate with other clients. Simultaneously, the method effectively integrates personal information with the social information generated by the global model from the perspective of Bayesian de- cision rules and evidence theory, enabling individuals to grow better with the help of collective wisdom. An inter- pretable federated learning mind is well scalable, and the experimental results indicate that the proposed method has superior robustness and accuracy than other state-of-the- art federated learning algorithms.

1. Introduction Federated learning is a new machine learning technique with various applications in data privacy protection and data security [19, 21, 37]. It can be viewed as social learning involving multiple agents coinciding with collective intelli- gence [3, 11, 13]. Unlike ordinary federated learning, per- sonalized federated learning can address the problem of data heterogeneity among clients and thereby improve their capabilities in relatively more realistic scenarios [4, 18]. However, designing a reliable and interpretable federated learning framework remains a significant challenge in the *Corresponding author. Figure 1. Cooperation between clients. Uninterpretable simple ag- gregation produces a global model that is not helpful for all clients because the information containing classes 2 and 3 may be nega- tive for client 4, as well as for clients 1 and 4. Clients 1 and 2 can well identify classes 1, 2, 3, and 4; however, unreliable random selection does not guarantee the participation of clients 1 and 2 in aggregation, whereas the simultaneous selection of less-capable clients such as clients 3 and 4 does. In this case, smart customers cannot offer more help to the not-so-smart ones. field of federated learning. FedProx [32], SCAFFOLD [15], MOON [17] used the global model to impose different con- straints on the client’s local training process. Consequently, the knowledge of the global model was better absorbed. Although [6, 33, 40] solved the problem of client het- erogeneity in a personalized federated learning framework by incorporating techniques such as clustering and knowl- edge distillation. [22, 35] propose a certain degree of inter- pretable client aggregation from the perspective of client contribution to the group. However, their selection and training of clients are often unreliable and uninterpretable, resulting in uncertainty in the training process and a ten- dency to ignore synergies between clients when the number of clients is large and the data distribution widely varies. Consequently, the collective intelligence is underutilized, as shown in Fig. 1. Herein, we propose a reliable and interpretable feder- ated learning framework, RIPFL. Specifically, we introduce Dempster–Shafer evidence theory (DST) [20, 34] to quan- 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. 20422 tify the uncertainty and performance of each client and pro- vide reliable client selection strategies. To reliably explain client choices and aggregation methods without wasting collective intelligence, RIPFL ensures that all smart clients participate in the aggregation process while a small num- ber of nonsmart individuals participate, to enable that nons- mart clients can adequately gain more valuable collective knowledge from smart clients. Moreover, a method that can reliably integrate social and personal information is pro- posed. The proposed framework is primarily applicable to situations where the number of clients is large and the tasks solved by clients are complex. The main contributions of this paper are as follows. • This paper proposes a reliable and interpretable per- sonalized FL architecture from the perspective of so- cial learning, which consists of interpretable local training, reliable clients selection and division, and ef- fective federated aggregation. • To reliably select the required clients, this paper intro- duces evidence theory to the local training of clients, thus quantifying the uncertainty of each client and pro- viding interpretable training methods. • A Bayesian-rule-based evidence fusion method is in- troduced by considering the global model as the prior information of clients when there are differences in the data distribution among clients. Consequently, the knowledge of the global model is not forgotten by clients in local training.

Pena_Re-Basin_via_Implicit_Sinkhorn_Differentiation_CVPR_2023

Abstract The recent emergence of new algorithms for permuting models into functionally equivalent regions of the solution space has shed some light on the complexity of error sur- faces and some promising properties like mode connectivity. However, finding the permutation that minimizes some ob- jectives is challenging, and current optimization techniques are not differentiable, which makes it difficult to integrate into a gradient-based optimization, and often leads to sub- optimal solutions. In this paper, we propose a Sinkhorn re- basin network with the ability to obtain the transportation plan that better suits a given objective. Unlike the cur- rent state-of-art, our method is differentiable and, there- fore, easy to adapt to any task within the deep learning do- main. Furthermore, we show the advantage of our re-basin method by proposing a new cost function that allows per- forming incremental learning by exploiting the linear mode connectivity property. The benefit of our method is com- pared against similar approaches from the literature un- der several conditions for both optimal transport and linear mode connectivity. The effectiveness of our continual learn- ing method based on re-basin is also shown for several com- mon benchmark datasets, providing experimental results that are competitive with the state-of-art. The source code is provided at https://github.com/fagp/sinkhorn-rebasin.

1. Introduction Despite the success of deep learning (DL) across many application domains, the loss surfaces of neural networks (NNs) are not well understood. Even for shallow NNs, the number of saddle points and local optima can increase expo- nentially with the number of parameters [4,13]. The permu- AB P(B) (a) Naive WM [2] Sinkhorn C() (b) Figure 1. (a) Loss landscape for the polynomial approximation task [27].AandBare models found by SGD. LMC suggests that re-basin the model Bwould result in a functionally equiv- alent model P(B), with no barrier on its linear interpolation (1)A+P(B). (b) Comparison of the cost in the lin- ear path along before and after re-basin using weight matching (WM) [2] and our Sinkhorn. The dashed line in the figures corre- sponds with the naive path, and the solid line is the path after the proposed Sinkhorn re-basin. The blue line represents WM path. 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. 20237 tation symmetry of neurons in each layer allows the same function to be represented with many different parameter values of the same network. Symmetries imposed by these invariances help us to better understand the structure of the loss landscape [6, 11, 13]. Previous studies establish that minima found by Stochas- tic Gradient Descent (SGD) are not only connected in the network parameter’s space by a path of non-increasing loss, but also permutation symmetries may allow us to con- nect those points linearly with no detriment to the loss [9, 11–13, 15, 24]. This phenomenon is often referred to as linear mode connectivity (LMC) [24]. For instance, Fig. 1a shows a portion of the loss landscape for the polynomial approximation task [27] using the method proposed by Li et al. [16].AandBare two minima found by SGD in different basins with an energy barrier between the pair. LMC suggests that if one considers permutation invariance, we can teleport solutions into a single basin where there is almost no loss barrier between different solutions [2, 11]. In literature, this mechanism is called re-basin [2]. How- ever, efficiently searching for permutation symmetries that bring all solutions to one basin is a challenging problem [11]. Three main approaches for matching units between two NNs have been explored in the literature [2]. Some studies propose a data-dependent algorithm that associates units across two NNs by matching their activations [2, 26]. Since activation-based matching is data dependent, it helps to adjust permutations to certain desired kinds of classes or domains [26]. Instead of associating units by their activa- tions, one could align the weights of the model itself [2,26], which is independent of the dataset, and therefore the com- putational cost is much lower. Finally, the third approach is to iteratively adjust the permutation of weights. In par- ticular, Ainsworth et al. [2] have proposed alternating be- tween models alignment and barrier minimization using a Straight-Through Estimator. Unfortunately, the proposed approaches so far are either non-differentiable [2, 11, 26] or computationally expensive [2], making the solution difficult to be extended to other applications, with a different objec- tive. For instance, adapting those methods for incremental learning by including the algorithm for weight matching be- tween two models trained on different domains is not trivial because of the difficulties in optimizing new objectives. In this work, inspired by [21], we relax the permutation matrix with the Sinkhorn operator [1], and use it to solve the re-basin problem in a differentiable fashion. To avoid the high cost for computing gradients in the proposal of Mena et al. [21], we use the implicit differentiation algo- rithm proposed in [10], which has been shown to be more cost-effective. Our re-basin formulation allows defining any differentiable objective as a loss function. A direct application of re-basin is the merger of diverse models without significantly degrading their performance[2, 5, 12, 13, 28]. Applications like federate learning [2], ensembling [12], or model initialization [5] exploit such a merger by selecting a model in the line connecting the mod- els to be combined. To show the effectiveness of our ap- proach, we propose a new continual learning algorithm that combines models trained on different domains. Our con- tinual learning algorithm differs from previous state-of-art approaches [22] because it directly estimates a model at the intersection of previous and new knowledge, by exploiting the LMC property observed in SGD-based solutions. Our main contribution can be summarized as follows: (1)Solving the re-basin for optimal transportation using implicit Sinkhorn differentiation, enabling better differen- tiable solutions that can be integrated on any loss. (2)A powerful way to use our re-basin method based on the Sinkhorn operator for incremental learning by considering it as a model merging problem and leveraging LMC. (3)An extensive set of experiments that validate our method for: (i) finding the optimal permutation to transform a model to another one equivalent; (ii) linear mode connec- tivity, to linearly connect two models such that their loss is almost identical along the entire connecting line in the weights space; and (iii) learning new domains and tasks in- crementally while not forgetting the previous ones.

Li_Unified_Mask_Embedding_and_Correspondence_Learning_for_Self-Supervised_Video_Segmentation_CVPR_2023

Abstract The objective of this paper is self-supervised learning of video object segmentation. We develop a unified framework which simultaneously models cross-frame dense correspon- dence for locally discriminative feature learning and embeds object-level context for target-mask decoding. As a result, it is able to directly learn to perform mask-guided sequential segmentation from unlabeled videos, in contrast to previous efforts usually relying on an oblique solution — cheaply “copying” labels according to pixel-wise correlations. Con- cretely, our algorithm alternates between i)clustering video pixels for creating pseudo segmentation labels ex nihilo; and ii)utilizing the pseudo labels to learn mask encoding and de- coding for VOS. Unsupervised correspondence learning is further incorporated into this self-taught, mask embedding scheme, so as to ensure the generic nature of the learnt repre- sentation and avoid cluster degeneracy. Our algorithm sets state-of-the-arts on two standard benchmarks ( i.e., DAVIS 17 and YouTube-VOS), narrowing the gap between self- and fully-supervised VOS, in terms of both performance and net- work architecture design.

1. Introduction In this article, we focus on a classic computer vision task: accurately segmenting desired object(s) in a video sequence, where the target object(s) are defined by pixel-wise mask(s) in the first frame. This task is referred as ( one-shot )video ob- ject segmentation (VOS) or mask propagation [1], playing a vital role in video editing and self-driving. Prevalent solu- tions [2–25] are built upon fully supervised learning techni- ques, costing intensive labeling efforts. In contrast, we aim to learn VOS from unlabeled videos — self-supervised VOS. Due to the absence of mask annotation during training, existing studies typically degrade such self-supervised yet mask-guided segmentation task as a combo of unsupervised correspondence learning and correspondence based, non- *Work done during an internship at Baidu VIS. †Corresponding author: Wenguan Wang . t+1 t+1 t t matching t t+1 mask warpingmask embeddingmatching mask decodingmask -guided segmentation (a) (b) (d) (c)ResNet-18 ResNet-50 MAST[28] VFS[39] LIIR[31] UVC[35] ConCor[41] Ours555963677175J&F 20 30 40 50 60 70 80 frame number4050607080 J&F UVC[35]ConCor[41]MAST[28]CRW[34]VFS[39]LIIR[31]Ours Figure 1. (a) Correspondence learning based self-supervised VOS, where mask tracking is simply degraded as correspondence match- ing mask warping. (b) We achieve self-supervised VOS by jointly learning mask embedding and correspondence matching. Our algo- rithm explicitly embeds masks for target object modeling, hence enabling mask-guided segmentation. (c) Performance comparison and (d) Performance over time, reported on DA VIS 17[42]val. learnable mask warping (cf.Fig. 1(a)). They first learn pixel- /patch-wise matching ( i.e., cross-frame correspondence) by exploring the inherent continuity in raw videos as free super- visory signals, in the form of i)aphotometric reconstruc- tionproblem where each pixel in a target frame is desired to be recovered by copying relevant pixels in reference frame(s) [26–31]; ii)acycle-consistency task that enforces matching of pixels/patches after forward-backward tracking [32–36]; andiii)acontrastive matching scheme that contrasts confi- dent correspondences against unreliable ones [37–40]. Once trained, the dense matching model is used to approach VOS in a cheap way (Fig.1(a)): the label of a query pixel/patch is simply borrowed from previously segmented ones, accord- ing to their appearance similarity (correspondence score). Though straightforward, these correspondence based “ex- pedient” solutions come with two severe limitations: First , 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. 18706 they learn to match pixels instead of customizing VOS tar- get – mask-guided segmentation, leaving a significant gap between the training goal and task/inference setup. During training, the model is optimized purely to discovery reliable, target-agnostic visual correlations, with no sense of object- mask information. Spontaneously, during testing/inference, the model struggles in employing first-/prior-frame masks to guide the prediction of succeeding frames. Second , from the view of mask-tracking, existing self-supervised solutions, in essence, adopt an obsolete, matching-/flow-based mask pro- pagation strategy [43–47]. As discussed even before the deep learning era[48–50], such a strategy is sub-optimal. Specifi- cally, without modeling the target objects, flow-based mask warping is sensitive to outliers, resulting in error accumula- tion over time [1]. Subject to the primitive matching-and- copy mechanism, even trivial errors are hard to be corrected, and oftenleadtomuchworseresultscausedbydriftsorocclu- sions. This is also why current top-leading fully supervised VOSsolutions[4, 5, 10–22]largely followa maskembedding learning philosophy — embedding frame-mask pairs , in- stead of only frame images, into the segmentation network. With such explicit modeling of the target object, more ro- bust and accurate mask-tracking can be achieved [1, 51]. Motivated by the aforementioned discussions, we inte- grate mask embedding learning and dense correspondence modeling into a compact, end-to-end framework for self- supervised VOS ( cf.Fig. 1(b)). This allows us to inject the mask-tracking nature of the task into the very heart of our algorithm and model training. However, bringing the idea of mask embedding into self-supervised VOS is not trivial, due to the lack of mask annotation. We therefore achieve mask embedding learning in a self-taught manner. Concretely, our model is trained by alternating between i)space-time pixel clustering, and ii)mask-embedded segmentation learning. Pixel clustering is to automatically discover spatiotempo- rally coherent object(-like) regions from raw videos. By uti- lizing such pixel-level video partitions as pseudo ground- truths of target objects, our model can learn how to extract target-specific context from frame-mask pairs, and how to leverage such high-level context to predict the next-frame mask. At the same time, such self-taught mask embedding scheme is consolidated by self-supervised dense correspon- dence learning. This allows our model to learn transferable, locally discriminative representations by making full use of the spatiotemporal coherence in natural videos, and prevent the degenerate solution of the deterministic clustering. Our approach owns a few distinctive features: First , it has the ability of directly learning to conduct mask-guided se- quential segmentation; its training objective is completely aligned with the core nature of VOS. Second , by learning to embed object-masks into mask tracking, target-oriented context can be efficiently mined and explicitly leveraged for object modeling, rather than existing methods merelyrelying on local appearance correlations for label “copy- ing”. Hence our approach can reduce error accumulation (cf.Fig. 1(d)) and perform more robust when the latent cor- respondences are ambiguous, e.g.,deformation, occlusion or one-to-many matches. Third , our mask embedding strategy endows our self-supervised framework with the potential of being empowered by more advanced VOS model designs developed in the fully-supervised learning setting. Through embracing the powerful idea of mask embed- ding learning as well as inheriting the merits of correspon- dence learning, our approach favorably outperforms state-of- the-art competitors, i.e.,3.2%,2.5%, and 2.2% mIoU gains on DA VIS 17[42]val, DA VIS 17test-dev and YouTube- VOS [52] val, respectively. In addition to narrowing the performance gap between self- and fully-supervised VOS, our approach establishes a tight coupling between them in the aspect of model design. We expect this work can foster the mutual collaboration between these two relevant fields.

Nakhli_Sparse_Multi-Modal_Graph_Transformer_With_Shared-Context_Processing_for_Representation_Learning_CVPR_2023

Abstract Processing giga-pixel whole slide histopathology images (WSI) is a computationally expensive task. Multiple in- stance learning (MIL) has become the conventional ap- proach to process WSIs, in which these images are split into smaller patches for further processing. However, MIL- based techniques ignore explicit information about the in- dividual cells within a patch. In this paper, by defining the novel concept of shared-context processing, we designed a multi-modal Graph Transformer (AMIGO) that uses the cel- lular graph within the tissue to provide a single representa- tion for a patient while taking advantage of the hierarchical structure of the tissue, enabling a dynamic focus between cell-level and tissue-level information. We benchmarked the performance of our model against multiple state-of-the-art methods in survival prediction and showed that ours can significantly outperform all of them including hierarchical Vision Transformer (ViT). More importantly, we show that our model is strongly robust to missing information to an extent that it can achieve the same performance with as low as 20% of the data. Finally, in two different cancer datasets, we demonstrated that our model was able to strat- ify the patients into low-risk and high-risk groups while other state-of-the-art methods failed to achieve this goal. We also publish a large dataset of immunohistochemistry images (InUIT) containing 1,600 tissue microarray (TMA) cores from 188 patients along with their survival informa- tion, making it one of the largest publicly available datasets in this context.

1. Introduction Digital processing of medical images has recently at- tracted significant attention in computer vision communi- ties, and the applications of deep learning models in this domain span across various image types ( e.g., histopathol- ogy images, CT scans, and MRI scans) and numerous tasks ( e.g., classification, segmentation, and survival pre- diction) [6, 11, 27, 28, 30, 36, 38, 44]. The paradigm-shifting ability of these models to learn predictive features directly from raw images has presented exciting opportunities in medical imaging. This has especially become more im- portant for digitized histopathology images where each data point is a multi-gigapixel image (also referred to as a Whole Slide Image or WSI). Unlike natural images, each WSI has high granularity at different levels of magnification and a size reaching 100,000 ×100,000 pixels, posing exciting challenges in computer vision. The typical approach to cope with the computational complexities of WSI processing is to use the Multiple In- stance Learning (MIL) technique [31]. More specifically, this approach divides each slide into smaller patches ( e.g., 256×256 pixels), passes them through a feature extractor, and represents the slide with an aggregation of these rep- resentations. This technique has shown promising results in a variety of tasks, including cancer subtype classifica- tion and survival prediction. However, it suffers from sev- eral major issues. Firstly, considering the high resolution of WSIs, even a non-overlapping 256 ×256 window gener- ates a huge number of patches. Therefore, the subsequent aggregation method of MIL has to perform either a simple pooling operation [3, 17] or a hierarchical aggregation 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. 11547 add more flexibility [6]. Nevertheless, the former limits the representative power of the aggregator drastically, and the latter requires a significant amount of computational power. Secondly, this approach is strongly dependent on the size of the dataset, which causes the over-fitting of the model in scenarios where a few data points ( e.g., hundreds) are avail- able. Lastly, despite the fact that cells are the main com- ponents of the tissue, the MIL approach primarily focuses on patches, which limits the resolution of the model to a snapshot of a population of cells rather than a single cell. Consequently, the final representation of the slide lacks the mutual interactions of individual cells. Multiple clinical studies have strongly established that the heterogeneity of the tissue has a crucial impact on the outcome of cancer [32, 46]. For instance, high levels of im- mune infiltration in the tumor stroma were shown to cor- relate with longer survival and positive therapy response in breast cancer patients [46]. Therefore, machine learn- ing methods for histopathology image analysis are required to account for tumor heterogeneity and cell-cell interac- tions. Nonetheless, the majority of the studies in this do- main deal with a single image highlighting cell nuclei (re- gardless of cell type) and extra cellular matrix. Recently, few studies have investigated pathology images where vari- ous cell types were identified using different protein mark- ers [25, 42]. However, they still utilized a single-modal ap- proach (i.e., one cell type in an image), ignoring the multi- modal context (i.e., several cell types within the tissue) of these images. In this study, we explore the application of graph neural networks (GNN) for the processing of cellular graphs (i.e., a graph constructed by connecting adjacent cells to each other) generated from histopathology images (Fig. 1). In particular, we are interested in the cellular graph because it gives us the opportunity to focus on cell-level informa- tion as well as their mutual interactions. By delivering an adaptable focus at different scales, from cell level to tissue level, such information allows the model to have a multi- scale view of the tissue, whereas MIL models concentrate on patches with a preset resolution and optical magnifica- tion. The availability of cell types and their spatial loca- tion helps the model to find regions of the tissue that have more importance for its representation ( e.g., tumor regions or immune cells infiltrating into tumor cells). In contrast to the expensive hierarchical pooling in MIL methods [6], the message-passing nature of GNNs offers an efficient ap- proach to process the vast scale of WSIs as a result of weight sharing across all the graph nodes. This approach also re- duces the need for a large number of WSIs during training as the number of parameters is reduced. In this work, we introduce a spArse MultI-modal Graph transfOrmer model (AMIGO) for the representation learn- ing of histopathology images by using cells as the main Figure 1. Cellular graph built from a 4,000×4,000pixel TMA core stained with Ki67 biomarker. Each red point demonstrates a cell that has a positive response to Ki67 while the blue points show cells that had a negative response to this biomarker. The highlighted patches show representative areas of the tissue where the spatial distribution of cells and the structure of the tissue are different. A typical MIL method cannot capture this heterogeneity as it does not take into account the location of the patches and lacks explicit information about the specific cells present within a patch. building blocks. Starting from the cell level, our model gradually propagates information to a larger neighborhood of cells, which inherently encodes the hierarchical structure of the tissues. More importantly, in contrast to other works, we approach this problem in a multi-modal manner, where we can get a broader understanding of the tissue structure, which is quite critical for context-based tasks such as sur- vival prediction. In particular, for a single patient, there can be multiple histopathology images available, each high- lighting cells of a certain type (by staining cells with spe- cific protein markers), and resulting in a separate cellular graph (Fig. 2). Therefore, using a multi-modal approach, we combine the cellular graphs of different modalities to- gether to obtain a unified representation for a single patient. This also affirms our stance regarding the importance of cell type and the distinction between different cellular connec- tivity types. Aside from achieving state-of-the-art results, we notice that, surprisingly, our multi-modal approach is strongly robust to missing information, and this enables us to perform more efficient training by relying on this recon- struction ability of the network. Our work advances the frontiers of MIL, Vision Transformer (ViT), and GNNs in multiple directions: • We introduce the first multi-modal cellular graph pro- cessing model that performs survival prediction based on the multi-modal histopathology images with shared contextual information. • Our model eliminates the critical barriers of MIL mod- 11548 Graph SAGE SAGE P ool SAGE P ool SAGE P oolGraph SAGE Graph SAGE ++PLM Instance A�en�on & Norm Mul�-Head A�en�on Add & Norm Feed Forward Add & NormGraph SAGE SAGE P ool SAGE P ool SAGE P oolGraph SAGE Graph SAGE ++PLM Instance A�en�on & NormGraph SAGE SAGE P ool SAGE P ool SAGE P oolGraph SAGE Graph SAGE ++PLM Instance A�en�on & NormShared Shared Shar ed Sharedc) b)a) Modality 1 (CD8 S tain) Modality 2 (CD20 S tain) Modality 3 (K56 S tain)Figure 2. Overview of our proposed method. a) The Cellular graphs are first extracted from histopathology images stained with different biomarkers ( e.g., CD8, CD20, and Ki67) and are fed into the encoder corresponding to their modality. The initial layer of encoders is shared, allowing further generalization, while the following layers pick up functionalities unique to each modality. The graphs at the top depict the hierarchical pooling mechanism of the model. b) The representations obtained from multiple graph instances in each modality are combined via a shared instance attention layer (shared-context processing), providing a single representation vector. c) A Transformer is used to merge the resultant vectors to create a patient-level embedding that will be used for downstream tasks such as survival prediction. els, enabling efficient training of multi-gigapixel im- ages on a single GPU and outperforming all the base- lines including ViT. It also implements the hierarchi- cal structure of Vision Transformer while keeping the number of parameters significantly lower during end- to-end training. • We also publish a large dataset of IHC images contain- ing1,600tissue microarray (TMA) cores from 188pa- tients along with their survival information, making it one of the largest datasets in this context.

Qin_MotionTrack_Learning_Robust_Short-Term_and_Long-Term_Motions_for_Multi-Object_Tracking_CVPR_2023

Abstract The main challenge of Multi-Object Tracking (MOT) lies in maintaining a continuous trajectory for each tar- get. Existing methods often learn reliable motion pat- terns to match the same target between adjacent frames and discriminative appearance features to re-identify the lost targets after a long period. However, the reliability of motion prediction and the discriminability of appear- ances can be easily hurt by dense crowds and extreme oc- clusions in the tracking process. In this paper, we pro- pose a simple yet effective multi-object tracker, i.e., Mo- tionTrack, which learns robust short-term and long-term motions in a unified framework to associate trajectories from a short to long range. For dense crowds, we design a novel Interaction Module to learn interaction-aware mo- tions from short-term trajectories, which can estimate the complex movement of each target. For extreme occlusions, we build a novel Refind Module to learn reliable long- term motions from the target’s history trajectory, which can link the interrupted trajectory with its correspond- ing detection. Our Interaction Module and Refind Mod- ule are embedded in the well-known tracking-by-detection paradigm, which can work in tandem to maintain superior performance. Extensive experimental results on MOT17 and MOT20 datasets demonstrate the superiority of our approach in challenging scenarios, and it achieves state- of-the-art performances at various MOT metrics. Code is available at https://github.com/qwomeng/MotionTrack.

1. Introduction Multi-Object Tracking (MOT) is a fundamental task in computer vision, which has a wide range of applications, †Co-first authors.∗Corresponding author. Target Tracked trajectory Occluded trajectory Video sequence (b) (a)Figure 1. Illustration of challenging scenarios in different videos . (a)Dense crowds. Pedestrians do not move independently in this situation. They will be affected by their surrounding neigh- bors to avoid collisions which will make their motion patterns hard to learn in practice. (b) Extreme occlusion. Pedestrians are eas- ily occluded by fixed facilities for a long period, such as billboard and sunshade, in which the dynamic environment will make them undergone a large appearance variation. such as autonomous driving [8] and intelligent surveil- lance [29]. It aims at jointly locating targets through bound- ing boxes and recognizing their identities throughout a whole video [40]. Though great progress has been made in the past few years, MOT still remains a challenging task due to the dynamic environment, such as dense crowds and extreme occlusions, in the tracking scenario. In general, the existing MOT methods either follow the tracking-by-detection [2] or tracking-by-regression [39, 40, 59], paradigm. The former methods first detect objects in each video frame and then associate detections between ad- jacent frames to create individual object tracks over time. The latter methods conduct tracking differently: the ob- ject detector not only provides frame-wise detections but also replaces the data association with a continuous regres- sion of each tracklet to its new position. Regardless of 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. 17939 paradigm, all methods need to address the short-range and long-range association problems, i.e., how to associate the alive tracklets with detections in a short time, and how to re-identify the lost tracklets with detections after a long pe- riod. For the short-range association problem, discriminative motion patterns and appearance features [5, 44] are of- ten learned to conduct data association between adjacent frames. However, as shown in Figure 1 (a), it is tough to learn discriminative representations in the dense crowd sce- nario. On the one hand, the bounding boxes of detections are too small to be distinguished by their appearances. On the other hand, different targets need to plan suitable paths to avoid collisions, which makes the resulting motions very complex in the tracking process. For the long-range asso- ciation problem, prior works [24, 43, 44] usually learn dis- criminative appearance features to re-identify the lost tar- gets after long occlusion [56–58]. As shown in Figure 1 (b), the main bottleneck of these methods is how to keep the ro- bustness of features against different poses, low resolution, and poor illumination for the same target. To alleviate this issue, the memory technology [7, 46] is widely applied to store diverse features for each target to match different tar- gets in a multi-query manner. Moreover, a lot of memo- ries and time will be consumed by the memory module and multi-query regime, unfriendly to real-time tracking. In this paper, we propose a simple yet effective object tracker, i.e., MotionTrack, to address the short-range and long-range association problems in MOT. In particular, our MotionTrack follows the tracking-by-detection paradigm, in which both interaction-aware and history trajectory- based motions are learned to associate trajectories from a short to long range. To deal with the short-range association problem, we design a novel Interaction Module to model all interactions between targets, which can predict their com- plex motions to avoid collisions. The Interaction Module uses an asymmetric adjacency matrix to represent the inter- action between targets, and obtains the prediction after the information fusion by a graph convolution network. Thanks to the captured target interaction, those short-term occluded targets can be successfully tracked in dense crowds. To deal with the long-range association problem, we design a novel Refind Module based on the history trajectory of each tar- get. It can effectively re-identify the lost targets through two steps: correlation calculation and error compensation. For the lost tracklets and the unmatched detections, the correla- tion calculation step takes the features of history trajectories and current detections as input, and computes a correlation matrix to represent the possibility that they are associated. Afterward, the error compensation step is further taken to revise the occluded trajectories. Extensive experiments on two benchmark datasets (MOT17 and MOT20) demonstrate that our proposed MotionTrack outperforms the previousstate-of-the-art methods. The main contribution of this work can be highlighted as follows: (1) We propose a simple yet effective multi-object tracker, MotionTrack, to address the short-range and long- range association problems; (2) We design a novel Interac- tion Module to model the interaction between targets, which can handle complex motions in dense crowds; (3) We design a novel Refind Module to learn discriminative motion pat- terns, which can re-identify the lost tracklets with current detections.

Mei_Exploring_and_Utilizing_Pattern_Imbalance_CVPR_2023

Abstract In this paper, we identify pattern imbalance from sev- eral aspects, and further develop a new training scheme to avert pattern preference as well as spurious correlation. In contrast to prior methods which are mostly concerned with category or domain granularity, ignoring the potential finer structure that existed in datasets, we give a new definition of seed category as an appropriate optimization unit to dis- tinguish different patterns in the same category or domain. Extensive experiments on domain generalization datasets of diverse scales demonstrate the effectiveness of the proposed method.

1. Introduction Over the past decade, the rise of deep neural networks (DNNs) has promoted the rapid development of various ar- tificial intelligence communities [13, 20, 22]. Despite the remarkable success, DNNs tend to take shortcuts to learn spurious features [24, 27]. The causal correlation between these spurious features and ground truth only exists in the training set, which hinders the generalization of DNN mod- els. This phenomenon is also known as domain shift. More- over, due to the incomplete distribution of training data, the learned model may have a preference for gender, race, and skin color, which will lead to serious ethical problems. To tackle these problems, various methods have been proposed to discuss the failure modes of out-of-distribution (OOD) generalization [18,30,32,43]. Some researchers fo- cus on encouraging the model to learn domain invariant fea- tures. Ganin et al. [9] simultaneously optimize a standard classifier and a domain classifier through adversarial train- ing, where the features extracted by DNN can be used for original classification but failed on domain recognition to inhibit domain characteristics learning. Arjovsky et al. [1] restrict the learned representations to be classified by sim- ilar optimal linear classifiers in different domains. Other researchers start by avoiding spurious features. Zhang et *Corresponding author.al. [42] argue that there exist sub-networks with preferable domain generalization ability in the model and represent the sub-network through a learnable mask. Nam et al. [28] as- sume that the spurious features are generally embodied in the texture or style of the image. They design SagNet to decouple the content and style of the image, impelling the feature extractor to pay more attention to the content infor- mation. Most of the above methods manually design spe- cific model structures to handle domain generalization. Instead of designing specific networks, we are more concerned about solving domain generalization by ex- ploring the character of the dataset. In particular, sup- pose a simple handwritten digit recognition scenario, where a large amount of digit 0 possesses the red background and digit 1 possesses the green background. The dataset with only the above two patterns cannot be effectively learned, since the model has no idea whether the task is to classify the digits or the background color. Therefore, in a given learnable data set, there must exist a minority of digit 0 with green background and digit 1 with red background. These samples play a significant role in establishing the true causal relationship between images and labels but have not been paid enough attention. We call pattern imbalance the phenomenon that different patterns in the same class ap- pear imbalanced, thus leading to model learning preference. Based on the above observations, we attribute the domain generalization problem to the mining of hard or minority patterns under imbalanced patterns. First of all, we iden- tify the pattern imbalance in the dataset from several per- spectives. We note that even though a model has achieved favorable performance on average, Achilles’heel still exists on some weak patterns. To alleviate the influence caused by imbalance patterns, we pay more attention to these samples of minority patterns and propose a training scheme based on dynamic distribution. To this end, we define a new concept, seed category, that is, the inherent pattern to distinguish, to promote model training by paying full attention to various patterns in the data set. Specifically, for samples of the same class, the seed category is divided based on the distance of the samples in the embedding space as a more fine-grained weight allocation unit than previous methods [19, 32, 39]. 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. 7569 In this paper, this dynamic and fine-grained training scheme enables our method to obtain excellent domain generaliza- tion performance. We argue that it is effective to apply more detailed weight allocation on out-of-distribution generalization tasks, that is, the patterns that are crucial but laborious to be learned by the model deserve special treatments, which is the most significant difference between our method and the previ- ous methods. Prior methods, e.g., GroupDRO [32], mini- mize the worst-case loss over domains to treat different do- mains differently, and the performance will be limited by the coarse granularity of grouped distribution. On the con- trary, the flexibility of our method is revealed in two as- pects, that is, the weight allocation unit is more detailed and the seed category can be dynamically adjusted during the training process. Our contributions can be summarized as follows: • We identify pattern imbalance generally existed in classification tasks and give a new definition of seed category, that is, the inherent pattern to recognize. • We further develop a dynamic weight distribution training strategy based on seed category to facilitate out-of-distribution performance. • Extensive experiments on several domain generaliza- tion datasets well demonstrate the effectiveness of the proposed method.

Pang_DPE_Disentanglement_of_Pose_and_Expression_for_General_Video_Portrait_CVPR_2023

Abstract One-shot video-driven talking face generation aims at producing a synthetic talking video by transferring the facial motion from a video to an arbitrary portrait image. Head pose and facial expression are always entangled in facial motion and transferred simultaneously. However, the entan- glement sets up a barrier for these methods to be used in video portrait editing directly, where it may require to modify the expression only while maintaining the pose unchanged. One challenge of decoupling pose and expression is the lack of paired data, such as the same pose but different expres- *Corresponding Authors.sions. Only a few methods attempt to tackle this challenge with the feat of 3D Morphable Models (3DMMs) for explicit disentanglement. But 3DMMs are not accurate enough to capture facial details due to the limited number of Blend- shapes, which has side effects on motion transfer. In this paper, we introduce a novel self-supervised disentanglement framework to decouple pose and expression without 3DMMs and paired data, which consists of a motion editing module, a pose generator, and an expression generator. The editing module projects faces into a latent space where pose motion and expression motion can be disentangled, and the pose or expression transfer can be performed in the latent space conveniently via addition. The two generators render 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. 427 modified latent codes to images, respectively. Moreover, to guarantee the disentanglement, we propose a bidirectional cyclic training strategy with well-designed constraints. Eval- uations demonstrate our method can control pose or expres- sion independently and be used for general video editing. Code: https://github.com/Carlyx/DPE

1. Introduction Talking face generation has seen tremendous progress in visual quality and accuracy over recent years. Literature can be categorized into two groups, i.e., audio-driven [23] and video-driven [16]. The former focuses on animating an unseen portrait image or video with a given audio. The latter aims at animating with a given video. Talking face generation has a variety of meaningful applications, such as digital human animation, film dubbing, etc. In this work, we target video-driven talking face generation. Recently, most methods [16, 26, 36, 39, 44] endeavor to drive a still portrait image with a video from different per- spectives, i.e., one-shot talking face generation. But only a few [19, 21, 30] make effort to reenact the portrait in a video with another talking video, i.e., video portrait editing. This is a more challenging task because edited faces are required to paste back to the original video and temporal dynamics need to be maintained. Several methods [19, 28] provide personal- ized solutions to this challenge by training a model on the videos of a specific person only. However, the learned model cannot generalize to other identities as the personalized train- ing heavily overfits the facial motion of the specific person and the background. For general video portrait editing, there- fore, resorting to the generalization property of one-shot talking face generation might be a feasible solution. One-shot methods can transfer facial motion from a driv- ing face to a source one, resulting in that the edited face mimics the head pose and facial expression *of the driving one. The facial motion consists of entangled pose motion and expression motion, which are always transferred simul- taneously in previous methods. However, the entanglement makes those methods unable to transfer pose or expression independently. Since the input to the processing network is always the cropped face rather than the full original image, if the pose is modified along with the expression, the paste- back operation can cause noticeable artifacts around the crop boundary, e.g., twisted neck and inconsistent background. Consequently, most one-shot methods face this obstacle pre- venting their application to general video portrait editing. One challenge to disentangle pose and expression is the lack of paired data, such as the same pose but different ex- pressions, or vice versa. In the literature, there are only a few exceptions that can get rid of this limitation, e.g., PIRen- *Note that facial expression here differs from emotion.derer [25] and StyleHEAT [41], which are based on 3D Morphable Models (3DMMs) [3], a predefined parametric representation that decomposes expression, pose, and iden- tity. However, the 3DMM-based methods heavily depend on the decoupling accuracy of the 3DMM parameters, which is far from satisfactory to reconstruct facial details due to the limited number of Blendshapes. Besides, optimization- based 3DMM parameter estimation is not efficient while learning-based estimation will introduce more errors. In this work, we propose a novel self-supervised dis- entanglement framework to decouple pose and expression, breaking through the limitation of paired data without us- ing 3DMMs. Our framework has a motion editing module, a pose generator, and an expression generator. The editing module projects faces into a latent space where coupled pose and expression motion in a latent code can be disentangled by a network. Then, pose or expression transfer can be per- formed by directly adding the latent code of a source face with the disentangled pose or expression motion code of a driving face. Finally, the two generators render modified latent codes to images. More importantly, to accomplish the disentanglement without paired data, we introduce a bidirec- tional cyclic training method with well-designed constraints. Specifically, given a source face Sand a driving one D, we transfer the expression and pose from DtoSsequentially, resulting in two synthetic faces, S′andS′′. Since there is no paired data, no supervision is provided for S′. To tackle the missing supervision, we exchange the role of SandD to transfer the pose and expression motion from StoD, resulting in D′andD′′. The distance between D′andS′is one constraint for learning. However, it is still not enough for disentangling pose and expression. Then, we discover another core constraint, i.e.,face reconstruction. When S andDare the same, S′andD′are exactly the same as Sand D, respectively. More analyses will be presented in Sec. 3. Our main contributions are three-fold: •We propose a self-supervised disentanglement frame- work to decouple pose and expression for independent motion transfer, without using 3DMMs and paired data. •We propose a bidirectional cyclic training strategy with well-designed constraints to achieve the disentangle- ment of pose and expression. •Extensive experiments demonstrate that our method can control pose or expression independently, and can be used for general video editing.

Li_PREIM3D_3D_Consistent_Precise_Image_Attribute_Editing_From_a_Single_CVPR_2023

Abstract We study the 3D-aware image attribute editing problem in this paper, which has wide applications in practice. Re- cent methods solved the problem by training a shared en- coder to map images into a 3D generator’s latent space or by per-image latent code optimization and then edited im- ages in the latent space. Despite their promising results near the input view, they still suffer from the 3D inconsis- tency of produced images at large camera poses and im- precise image attribute editing, like affecting unspecified attributes during editing. For more efficient image inver- sion, we train a shared encoder for all images. To alle- viate 3D inconsistency at large camera poses, we propose two novel methods, an alternating training scheme and a multi-view identity loss, to maintain 3D consistency and subject identity. As for imprecise image editing, we at- tribute the problem to the gap between the latent space of real images and that of generated images. We compare the latent space and inversion manifold of GAN models and demonstrate that editing in the inversion manifold can achieve better results in both quantitative and qualitative *Corresponding authors.evaluations. Extensive experiments show that our method produces more 3D consistent images and achieves more precise image editing than previous work. Source code and pretrained models can be found on our project page: https://mybabyyh.github.io/Preim3D/ .

1. Introduction Benefiting from the well-disentangled latent space of Generative Adversarial Networks (GANs) [12], many works study GAN inversion [1, 2, 11, 28, 35, 36, 40] as well as real image editing in the latent space [14, 15, 22, 31, 32]. With the popularity of Neural Radiance Fields (NeRF) [24], some works start to incorporate it into GAN frameworks for unconditional 3D-aware image generation [6, 7, 13, 25–27, 30]. In particular, EG3D [6], the state-of-the-art 3D GAN, is able to generate high-resolution multi-view-consistent images and high-quality geometry conditioned on gaus- sian noise and camera pose. Similar to 2D GANs, 3D GANs also have a well semantically disentangled latent space [6, 13, 21, 33], which enables realistic yet challeng- ing 3D-aware image editing. Achieving 3D-aware image editing is much more chal- 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. 8549 lenging because it not only has to be consistent with the input image at the input camera pose but also needs to pro- duce 3D consistent novel views. Recently, 3D-Inv [21] uses pivotal tuning inversion (PTI) [29], first finding out a piv- otal latent code and then finetuning the generator with the fixed pivotal latent code, to obtain the latent code and edit the image attributes in the latent space. IDE-3D [33] pro- poses a hybrid 3D GAN inversion approach combining tex- ture and semantic encoders and PTI technique, accelerating the optimization process by the encoded initial latent code. Pixel2NeRF [5] is the first to achieve 3D inversion by train- ing an encoder mapping a real image to the latent space Z ofπ-GAN [7]. However, these methods still do not solve the problem of 3D consistency at large camera poses and precise image attribute editing. As shown in Fig. 4, some inverted images meet head distortion at large camera poses, or some unspecific attributes of edited images are modified. In this paper, we propose a pipeline that enables PRecise Editing in the Inversion Manifold with 3Dconsistency effi- ciently, termed PREIM3D . There are three goals to achieve for our framework, (i) image editing efficiently , (ii) precise inversion , which aims to maintain realism and 3D consis- tency of multiple views, and (iii) precise editing , which is to edit the desired attribute while keeping the other attributes unchanged. 3D-Inv and IDE-3D optimized a latent code for each image, which is not suitable for interactive applica- tions. Following Pixel2NeRF, we train a shared encoder for all images for efficiency. To address precise inversion , we introduce a 3D consis- tent encoder to map a real image into the latent space W+ of EG3D, and it can infer a latent code with a single forward pass. We first design a training scheme with alternating in- domain images (i.e., the generated images) and out-domain images (i.e., the real images) to help the encoder maintain the 3D consistency of the generator. We optimize the en- coder to reconstruct the input images in the out-domain im- age round. In the in-domain image round, we additionally optimize the encoder to reconstruct the ground latent code, which will encourage the distribution of the inverted latent code closer to the distribution of the original latent code of the generator. Second, to preserve the subject’s identity, we propose a multi-view identity loss calculated between the input image and novel views randomly sampled in the sur- rounding of the input camera pose. Though many works tried to improve the editing pre- cision by modifying latent codes in Zspace [31], W space [14, 17, 34], W+space [1, 1, 40], and Sspace [37], they all still suffer from a gap between real image editing and generated image editing because of using the editing di- rections found in the original generative latent space to edit the real images. To bridge this gap, we propose a real im- age editing subspace, which we refer to inversion manifold . We compare the inversion manifold and the original latentspace and find the distortion between the attribute editing directions. We show that the editing direction found in the inversion manifold can control the attributes of the real im- ages more precisely. To our knowledge, we are the first to perform latent code manipulation in the inversion mani- fold. Our methodology is orthogonal to some existing edit- ing methods and can improve the performance of manipu- lation in qualitative and quantitative results when integrated with them. Figure 1 shows the inversion and editing results produced by our method. Given a single real image, we achieve 3D reconstruction and precise multi-view attribute editing. The contributions of our work can be summarized as fol- lows: • We present an efficient image attribute editing method by training an image-shared encoder for 3D-aware generated models in this paper. To keep 3D consis- tency at large camera poses, we propose two novel methods, an alternating training scheme and a multi- view identity loss, to maintain 3D consistency and sub- ject identity. • We compare the latent space and inversion manifold of GAN models, and demonstrate that editing in the inversion manifold can achieve better results in both quantitative and qualitative evaluations. The proposed editing space helps to close the gap between real image editing and generated image editing. • We conduct extensive experiments, including both quantitative and qualitative, on several datasets to show the effectiveness of our methods.

Muglikar_Event-Based_Shape_From_Polarization_CVPR_2023

Abstract State-of-the-art solutions for Shape-from-Polarization (SfP) suffer from a speed-resolution tradeoff: they either sacrifice the number of polarization angles measured or necessitate lengthy acquisition times due to framerate con- straints, thus compromising either accuracy or latency. We tackle this tradeoff using event cameras. Event cameras operate at microseconds resolution with negligible motion blur, and output a continuous stream of events that precisely measures how light changes over time asynchronously. We propose a setup that consists of a linear polarizer rotating at high speeds in front of an event camera. Our method uses the continuous event stream caused by the rotation to reconstruct relative intensities at multiple polarizer an- gles. Experiments demonstrate that our method outper- forms physics-based baselines using frames, reducing the MAE by 25% in synthetic and real-world datasets. In the real world, we observe, however, that the challenging con- ditions (i.e., when few events are generated) harm the per- formance of physics-based solutions. To overcome this, we propose a learning-based approach that learns to esti- mate surface normals even at low event-rates, improving the physics-based approach by 52% on the real world dataset. The proposed system achieves an acquisition speed equiva- lent to 50fps (>twice the framerate of the commercial po- larization sensor) while retaining the spatial resolution of 1MP . Our evaluation is based on the first large-scale dataset for event-based SfP . Code, dataset and video are available under: https://rpg.ifi.uzh.ch/esfp.html https://youtu.be/sF3Ue2Zkpec

1. Introduction Polarization cues have been used in many applications across computer vision, including image dehazing [41], panorama stitching and mosaicing [42], reflection removal [21], image segmentation [25], optical flow gyroscope, [47] and material classification [5]. Among these, Shape-from- Polarization (SfP) methods exploit changes in polariza- a) Setup Objectrotating polarizer ω timeevent streamevent camera t=τ unpolarized light b) Method Events att=τ Learning Based encoder decoderPhysics Based tSurface Normals (Ours) c) Baselines Images 135 deg 90 deg 45 deg 0 degMahmoud (2012) Ba (2020) Figure 1. Surface normal estimation using event-based SfP. (a) Rotating a polarizer in front of an event camera creates sinosoidal changes in intensities, triggering events. (b) The proposed event- based method uses the continuous event stream to reconstruct rel- ative intensities at multiple polarizer angles which is used to es- timate surface normals using physics-based and learning-based method. (c) Our approach outperforms image-based baselines [24, 51]. tion information to infer geometric properties of an object [2,18,22,49,51]. It uses variations in radiance under differ- ent polarizer angles to estimate the 3D surface of a given object. In particular, when unpolarized light is reflected from a surface, it becomes partially polarized depending on the geometry and material of the surface. Surface normals, and thus 3D shape, can then be estimated by orienting a po- larizing filter in front of a camera sensor and studying the relationship between the polarizer angle and the magnitude of light transmission. SfP has a number of advantages over both active and passive depth sensing methods. Unlike ac- 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. 1547 (a) Division of fo- cal plane (b) Division of time Figure 2. Illustration of SfP methods. tive depth sensors that use structured light (SL) [7, 48] or time-of-flight (ToF), SfP is not limited by material type and can be applied to non-Lambertian surfaces like transparent glass and reflective, metallic surfaces. Despite these advantages, however, estimating high- quality surface normals from polarization images is still an open challenge. Division of Focal Plane (DoFP) meth- ods [22, 32, 33, 51] trade-off spatial resolution for latency and allow for the capture of four polarizations in the same image. This is achieved through a complex manufactur- ing process that requires precisely placing a micro-array of four polarization filters on the image sensor [32, 33], as shown in Fig. 2a Despite the reduced latency, this system constrains the maximum number of polarization angles that can be captured, potentially impacting the accuracy of the estimates as we show in our results.Additionally, the spa- tial resolution of the sensor is also reduced, requiring fur- ther mosaicing-based algorithms for high-resolution recon- struction [50]. On the other hand, Division of Time (DoT) methods [2, 18, 49] provide full-resolution images and are not limited in the number of polarization angles they can capture thanks to a rotating polarizing filter put in front of the image sensor. The frame rate of the sensing camera, however, effectively limits the rate at which the filter can rotate, increasing the acquisition time significantly (acqui- sition time =N/f , where Nis the number of polarizer an- gles and fis the framerate of the camera). For this reason, commercial solutions, such as the Lucid Polarisens [33], fa- vor DoFP, despite the lower resolution of both polarization angles and image pixels. To overcome this shortcoming, re- cently, significant progress has been made with data-driven priors [22, 51]. However, these solutions still fall short in terms of computational complexity when compared to DoT methods. A solution able to bridge the accuracy of DoT with the speed of DoFP is thus still lacking in the field. In this paper, we tackle the speed-resolution trade-off using event cameras. Event cameras are efficient high- speed vision sensors that asynchronously measure changes in brightness intensity with microsecond resolution. We ex- ploit these characteristics to design a DoT approach able to operate at high acquisition speeds (up to 5,000fps vs. 22fps of standard frame-based devices) and full-resolution (1280×720) s shown in Fig. 1. Thanks to the working principles of event-cameras, our sensing device provides a continuous stream of information for estimating the sur-Dataset Modality (Resolution) Size Polar3D [18] 6 Images(DoT) 18 MP 3 DeepSfP [51] 4 Images(DoFP) 1224×1024 236 SPW [22] 4 Images(DoFP) 1224×1024 522 ESfP- Synthetic (Ours) Events (DoT) + 12 Images(DoT) 512×512 104 ESfP- Real (Ours) Events (DoT) + 4 Images (DoFP) 1280×720 90 Table 1. Summary of publicly available datasets for SfP. face normal as compared to the discrete intensities cap- tured at fixed polarization angles of traditional approaches. We present two algorithms to estimate surface normals from events, one using geometry and the other based on a learning-based approach. Our geometry-based method takes advantage of the continuous event stream to recon- struct relative intensities at multiple polarizer angles, which are then used to estimate the surface normal using tradi- tional methods. Since events provide a temporally rich in- formation, this results in better reconstruction of intermedi- ate intensities. This leads to an improvement of upto 25% in surface normal estimation, both on the synthetic dataset and on the real-world dataset.On the real dataset, however, the non-idealities of the event camera introduce a lower fill-rate (percentage of pixels triggering events) of 3.6%in average (refer Section 3.1). To overcome this, we propose a deep learning framework which uses a simple U-Net network to predict the dense surface normals from events. Our data- driven approach improves the accuracy over the geometry- based method by 52%. Our contributions can be summa- rized as follows: • A novel approach for shape-from-polarization using an event camera. Our approach utilizes the rich temporal information of events to reconstruct event intensities at multiple polarization angles. These event intensities are then used to estimate the surface normal. Our method outperforms previous state-of-the-art physics-based ap- proaches using images by 25% in terms of accuracy. • A learning-based framework which predicts surface nor- mals using events to solve the issue of low fill-rate com- mon in the real-world. This framework improves the es- timation over physics-based approach by 52% in terms of angular error. • Lastly, we present the firstlarge scale dataset containing over 90 challenging scenes for SfP with events and im- ages. Our dataset consists of events captured by rotating a polarizer in front of an event camera, as well as images captured using the Lucid Polarisens [33].

Lu_Robust_and_Scalable_Gaussian_Process_Regression_and_Its_Applications_CVPR_2023

Abstract This paper introduces a robust and scalable Gaussian process regression (GPR) model via variational learning. This enables the application of Gaussian processes to a wide range of real data, which are often large-scale and contaminated by outliers. Towards this end, we employ a mixture likelihood model where outliers are assumed to be sampled from a uniform distribution. We next derive a variational formulation that jointly infers the mode of data, i.e., inlier or outlier, as well as hyperparameters by maximizing a lower bound of the true log marginal like- lihood. Compared to previous robust GPR, our formula- tion approximates the exact posterior distribution. The in- ducing variable approximation and stochastic variational inference are further introduced to our variational frame- work, extending our model to large-scale data. We ap- ply our model to two challenging real-world applications, namely feature matching and dense gene expression impu- tation. Extensive experiments demonstrate the superiority of our model in terms of robustness and speed. Notably, when matching 4k feature points, its inference is completed in milliseconds with almost no false matches. The code is at github.com/YifanLu2000/Robust-Scalable-GPR .

1. Introduction Gaussian processes (GPs) [31] are probably the primary non-parametric method for inference on latent functions. They have a wide range of applications from biology [3] to computer vision [41]. A commonly used observation model for Gaussian process regression (GPR) is the Normal dis- tribution, which brings great convenience to the inference. Unfortunately, a well-known limitation of the Gaussian ob- servation model is its sensitivity to outliers in data. As il- lustrated in Fig. 1 (b), a few outliers can drastically destroy the entire posterior regression result. This hinders the real- world applications of GPR for many domains, where out- liers are often inevitable. This paper intends to conquer the GPR with outlier contaminated data. The idea of robust regression is not new. Outlier detec- *Corresponding Author Figure 1. Regression with our model. (a) Perform exact GPR from 100 inliers. (b) When there are only 6 outliers in the data, the exact GPR leads to completely wrong results. (c) By comparison, our model is able to recover the exact posterior even facing 100 outliers. (d) The feature matching result using our model. (e) The dense spatial gene expression imputation result using our model. tion has been extensively and systematically described in [6,9,10,29]. In the context of GPR, many efforts tried to re- place the Gaussian likelihood with other distributions show- ing heavy-tail behaviors, including Student- t[16,21,28,30], Laplace [22, 30], Gaussian mixture [8, 22, 27], and data- dependent noise model [17]. The challenge with these non- Gaussian likelihoods lies in the inference, which is analyti- cally intractable. To this end, many approximation schemes have been applied, despite having high computational com- plexity, e.g., Markov Chain Monte Carlo (MCMC) sam- pling and Expectation Propagation (EP) [22]. In this paper, we propose a more effective mixture like- lihood model, where uniform distribution accounts for the outliers and Gaussian for inliers. In our formulation, the outliers are independent of the GP and do not affect the computation of the posterior GP, thereby allowing to tol- erate more outliers. We next introduce a variational method that jointly determines the modes of data ( i.e., inlier or out- lier) as well as hyperparameters by maximizing a lower bound to the marginal likelihood. We highlight that the dif- ference between our variational formulation and pervious methods is that the modes of data now become variational parameters and are obtained by minimizing the Kullback- 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. 21950 Leibler (KL) divergence between the variational and the true posterior distribution. Thus, the proposed formulation is less likely to overfit and is able to approximate the exact posterior GP only from inliers, as in Fig. 1 (c). Inspired by [37], the sparse inducing variable approxi- mation is integrated into our variational framework, which retains the exact GP prior but performs posterior approx- imation, and reduces the time complexity from O(n3)to O(nm2). By treating the inducing variables as global vari- ables [18], our variational model enjoys the acceleration by Stochastic Variational Inference (SVI) [20]. It performs stochastic optimization from natural gradient and further decreases the time complexity to O(km2). This provides a guarantee for our model to scale to large-scale data. We apply our robust GPR model to two real-world ap- plications, say feature matching and dense spatial gene ex- pression imputation, as illustrated in the Figs. 1 (d) and (e). Extensive experiments demonstrate the superiority of our method on both numerical data and real applications. To summarize, our contributions include the following. (i) We present a robust Gaussian process regression model, which uses variational learning to approximate the true ex- act posterior. (ii) We leverage inducing variables and SVI to adapt our model to large-scale data. (iii) Two applica- tions of our model are described. Extensive experimental validation demonstrates the superiority of our model.

Potje_Enhancing_Deformable_Local_Features_by_Jointly_Learning_To_Detect_and_CVPR_2023

Abstract Local feature extraction is a standard approach in com- puter vision for tackling important tasks such as image matching and retrieval. The core assumption of most meth- ods is that images undergo affine transformations, disregard- ing more complicated effects such as non-rigid deformations. Furthermore, incipient works tailored for non-rigid corre- spondence still rely on keypoint detectors designed for rigid transformations, hindering performance due to the limita- tions of the detector. We propose DALF (Deformation-Aware Local Features), a novel deformation-aware network for jointly detecting and describing keypoints, to handle the challenging problem of matching deformable surfaces. All network components work cooperatively through a feature fusion approach that enforces the descriptors’ distinctiveness and invariance. Experiments using real deforming objects showcase the superiority of our method, where it delivers 8% improvement in matching scores compared to the previous best results. Our approach also enhances the performance of two real-world applications: deformable object retrieval and non-rigid 3D surface registration. Code for training, in- ference, and applications are publicly available at verlab. dcc.ufmg.br/descriptors/dalf_cvpr23 .

1. Introduction Finding pixel-wise correspondences between images de- picting the same surface is a long-standing problem in com- puter vision. Besides varying illumination, viewpoint, and distance to the object of interest, real-world scenes impose additional challenges. The vast majority of the correspon- dence algorithms in the literature assume that our world is rigid, but this assumption is far from the truth. It is notice- able that the community invests significant efforts into novel architectures and training strategies to improve image match- DALF DISKMS = 0.48 MS = 0.38Figure 1. Image matching under deformations . We propose DALF, a deformation-aware keypoint detector and descriptor for matching deformable surfaces. DALF (top) enables local feature matching across deformable scenes with improved matching scores (MS) compared to state-of-the-art, as illustrated with DISK [37]. Green lines show correct matches, and red markers, the mismatches. ing for rigid scenes [6, 19, 26, 34, 37, 42], but disregards the fact that many objects in the real world can deform in more complex ways than an affine transformation. Many applications in industry, medicine, and agricul- ture require tracking, retrieval, and monitoring of arbitrary deformable objects and surfaces, where a general-purpose matching algorithm is needed to achieve accurate results. Since the performance of standard affine local features sig- nificantly decreases for scenarios such as strong illumination changes and deformations, a few works considering a wider class of transformations have been proposed [24, 25, 30]. However, all the deformation-aware methods neglect the keypoint detection phase, limiting their applicability in chal- lenging deformations. Although the problems of keypoint detection and description can be treated separately, recent works that jointly perform detection and description of fea- tures [4, 26] indicate an entanglement of the two tasks since the keypoint detection can impact the performance of the de- scriptor. The descriptor for its turn can be used to determine reliable points optimized for specific goals. In this work, 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. 1306 we propose a new method for jointly learning keypoints and descriptors robust to deformations, viewpoint, and illumina- tion changes. We show that the detection phase is critical to obtain robust matching under deformations. Fig. 1 de- picts an image pair with challenging deformations, where our method can extract reliable keypoints and match them correctly, significantly increasing matching scores compared to the recent state-of-the-art approach DISK [37]. Contributions. (1) Our first contribution is a new end-to- end method called DALF (Deformation-Aware Local Fea- tures), which jointly learns to detect keypoints and extract descriptors with a mutual assistance strategy to handle sig- nificant non-rigid deformations. Our method boosts the state-of-the-art in this type of feature matching by 8% using only synthetic warps as supervision, showing strong gener- alization capabilities. We leverage a reinforcement learning algorithm for unified training, combined with spatial trans- formers that capture deformations by learning context priors affecting the image; (2)Second, we introduce a feature fu- sion approach, a major difference from previous methods that allows the model to tackle challenging deformations with complementary features (with distinctiveness and in- variance properties) obtained from both the backbone and the spatial transformer module. This approach is shown beneficial with substantial performance improvements com- pared to the non-fused features; (3)Finally, we demonstrate state-of-the-art results in non-rigid local feature applications for deformable object retrieval and non-rigid 3D surface reg- istration. We also will make the code and both applications publicly available to the community.

Ni_PATS_Patch_Area_Transportation_With_Subdivision_for_Local_Feature_Matching_CVPR_2023

Abstract Local feature matching aims at establishing sparse cor- respondences between a pair of images. Recently, detector- free methods present generally better performance but are not satisfactory in image pairs with large scale differences. In this paper, we propose Patch Area Transportation with Subdivision (PATS) to tackle this issue. Instead of build- ing an expensive image pyramid, we start by splitting the original image pair into equal-sized patches and gradu- ally resizing and subdividing them into smaller patches with the same scale. However, estimating scale differences be- tween these patches is non-trivial since the scale differ- ences are determined by both relative camera poses and scene structures, and thus spatially varying over image pairs. Moreover, it is hard to obtain the ground truth for real scenes. To this end, we propose patch area transporta- tion, which enables learning scale differences in a self- supervised manner. In contrast to bipartite graph match- ing, which only handles one-to-one matching, our patch area transportation can deal with many-to-many relation- ships. PATS improves both matching accuracy and cover- age, and shows superior performance in downstream tasks, such as relative pose estimation, visual localization, and optical flow estimation. The source code is available at https://zju3dv.github.io/pats/ .

1. Introduction Local feature matching between images is essential in many computer vision tasks which aim to establish corre- spondences between a pair of images. In the past decades, local feature matching [3, 40] has been widely used in a large number of applications such as structure from mo- tion (SfM) [44, 64], simultaneous localization and mapping (SLAM) [30,36,62], visual localization [19,41], object pose estimation [22, 61], etc. The viewpoint change from the *Junjie Ni and Yijin Li contributed equally to this work. †Guofeng Zhang is the corresponding author. Figure 1. Two-view reconstruction results of LoFTR [49], ASpanFormer [7], PDC-Net+ [58] and our approach on MegaDepth dataset [27]. PATS can extract high-quality matches under severe scale variations and in indistinctive regions with repetitive patterns, which allows semi-dense two-view reconstruc- tion by simply triangulating the matches in a image pair. In con- trast, other methods either obtain fewer matches or even obtain erroneous results. source image to the target image may lead to scale varia- tions, which is a long-standing challenge in local feature matching. Large variations in scale leads to two severe consequences: Firstly, the appearance is seriously distorted, which makes learning the feature similarity more challeng- ing and impacts the correspondence accuracy. Secondly, there may be several pixels in the source image correspond- ing to pixels in a local window in the target image. How- ever, existing methods [33, 40] only permit one potential target feature to be matched in the local window, and the following bipartite graph matching only allows one source pixel to win the matching. The coverage of correspondences derived from such feature matches is largely suppressed and will impact the downstream tasks. Before the deep learning era, SIFT [33] is a milestone that tackles the scale problem by detecting local features on an image pyramid and then matching features crossing pyramid levels, called scale-space analysis. This technique is also adopted in the inference stage of learning-based lo- cal features [39]. Recently, LoFTR abandons feature detec- tion stage and learns to directly draw feature matches via simultaneously encoding features from both images based 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. 17776 Figure 2. Scale Alignment with Patch Area Transportation. Our approach learns to find the many-to-many relationship and scale differences through solving the patch area transportation. Then we crop the patches and resize the image content to align the scale, which remove the appearance distortion. on the attention mechanism. By removing the information bottleneck caused by detectors, LoFTR [49] produces bet- ter feature matches. However, LoFTR does not handle the scale problem and the scale-space analysis is infeasible in this paradigm because conducting attention intra- and inter- different scales will bring unbearable increasing computa- tional costs. As a result, the scale curse comes back again. In this paper, we propose Patch AreaTransportation with Subdivision (PATS) to tackle the scale problem in a detector-free manner. The appearance distortion can be al- leviated if the image contents are aligned according to their scale differences before feature extraction. As shown in Fig. 2, if the target image is simply obtained by magnify- ing the source image twice, a siamese feature encoder will produce features with large discrepancies at corresponding locations. The discrepancies are corrected if we estimate the scale difference and resize the target image to half be- fore feature extraction. Considering that scale differences are spatially varying, we split the source image into equal- sized patches and then align the scale patch-wisely. Specif- ically, we identify a corresponding rectangular region in the target image for each source patch and then resize the im- age content in the region to align the scale. By also splitting the target image into patches, the rectangular regions can be represented with patches bounded by boxes. Based on this representation, one source patch corresponds to mul- tiple target patches. Moreover, the bounding box may be overlapped, indicating that one target patch may also corre- spond to multiple source patches. Here comes the question: how can we find many-to-many patch matches instead of one-to-one [7, 42, 49]? We observe that finding target patches for a source patch can be regarded as transporting the source patch to the target bounding box, where each target patch inside the box occupies a portion of the content. In other words, the area proportion that the target patches occupying the source patch should be summed to 1. Motivated by this observation, we propose to predict the target patches’ area and formulate patch matching as a patch area transporta- tion problem that transports areas of source patches to tar- get patches with visual similarity restrictions. Solving this problem with Sinkhorn [10], a differential optimal trans- port solver, also encourages our neural network to bettercapture complex visual priors. Once the patch matching is finished, the corresponding bounding boxes can be eas- ily determined. According to the patch area transportation with patch subdivision from coarse to fine, PATS signifi- cantly alleviates appearance distortion, which largely eases the difficulty of feature learning to measure visual similar- ity. Moreover, source patches being allowed to match over- lapped target patches naturally avoid the coverage reduction problem. After resizing the target regions according to es- timated scale differences, we subdivide the corresponding source patch and target region to obtain finer correspon- dences, dubbed as scale-adaptive patch subdivision. Fig. 1 shows qualitative results of our approach. Our contributions in this work can be summarized as three folds: 1) We propose patch area transportation to handle the many-to-many patch-matching challenge and grants the ability that learning scale differences in a self- supervised manner to the neural network. 2) We pro- pose a scale-adaptive patch subdivision to effectively re- fine the correspondence quality from coarse to fine. 3) Our patch area transportation with subdivision (PATS) achieves state-of-the-art performance and presents strong robustness against scale variations.

Peters_pCON_Polarimetric_Coordinate_Networks_for_Neural_Scene_Representations_CVPR_2023

Abstract Neural scene representations have achieved great suc- cess in parameterizing and reconstructing images, but cur- rent state of the art models are not optimized with the preservation of physical quantities in mind. While current architectures can reconstruct color images correctly, they create artifacts when trying to fit maps of polar quanti- ties. We propose polarimetric coordinate networks (pCON), a new model architecture for neural scene representations aimed at preserving polarimetric information while accu- rately parameterizing the scene. Our model removes arti- facts created by current coordinate network architectures when reconstructing three polarimetric quantities of inter- est. All code and data can be found at this link: https: //visual.ee.ucla.edu/pcon.htm .

1. Introduction Neural scene representations are a popular and useful tool in many computer vision tasks, but these models are optimized to preserve visual content, not physical informa- tion. Current state-of-the-art models create artifacts due to the presence of a large range of spatial frequencies when re- constructing polarimetric data. Many tasks in polarimetric imaging rely on precise measurements, and thus even small artifacts are a hindrance for downstream tasks that would like to leverage neural reconstructions of polarization im- ages. In this work we present pCON, a new architecture for neural scene representations. pCON leverages images’ sin- gular value decompositions to effectively allocate network capacity to learning the more difficult spatial frequencies at each pixel. Our model reconstructs polarimetric images without the artifacts introduced by state-of-the-art models. The polarization of light passing through a scene con- tains a wealth of information, and while current neural rep- resentations can represent single images accurately, but they produce noticeable visual artifacts when trying to represent *Equal contribution.multiple polarimetric quantities concurrently. We propose a new architecture for neural scene repre- sentations that can effectively reconstruct polarimetric im- ages without artifacts. Our model reconstructs color images accurately while also ensuring the quality of three impor- tant polarimetric quantities, the degree ( ρ) and angle ( ϕ)of linear polarization (DoLP and AoLP), and the unpolarized intensity Iun. This information is generally captured using images of a scene taken through linear polarizing filters at four different angles. Instead of learning a representation of these images, our model operates directly on the DoLP, AoLP and unpolarized intensity maps. When learning to fit these images, current coordinate network architectures produce artifacts in the predicted DoLP and unpolarized in- tensity maps. To alleviate this issue, we take inspiration from traditional image compression techniques and fit im- ages using their singular value decompositions. Images can be compressed by reconstructing them using only a subset of their singular values [28]. By utilizing different, non- overlapping sets of singular values to reconstruct an image, the original image can be recovered by summing the indi- vidual reconstructions together. Our model is supervised in a coarse-to-fine manner, which helps the model to represent both the low and and high frequency details present in maps of polarimetric quantities without introducint noise or tiling artifacts. A demonstration of the efficacy our model can be seen in Fig. 1 and Table 1. Furthermore, our model is capa- ble of representing images at varying levels of detail, creat- ing a tradeoff between performance and model size without retraining. 1.1. Contributions To summarize, the contributions of our work include: • a coordinate network architecture for neural scene rep- resentations of polarimetric images; • a training strategy for our network which learns a se- ries of representations using different sets of singular values, allowing for a trade-off between performance and model size without retraining; 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. 16579 GT SIREN [52] ACORN [34] ReLU P.E Ours Figure 1. Our model reconstructs the training scene more accurately than other architectures. Our model does not have the noise pattern present in reconstructions from SIREN [52] or a ReLU MLP with positional encoding [38], nor does it show tiling artifacts as in ACORN’s [34] prediction. • results demonstrating that our model reconstructs maps of polarimetric quantities without the artifacts created by current state-of-the-art approaches.

Qu_How_To_Prevent_the_Poor_Performance_Clients_for_Personalized_Federated_CVPR_2023

Abstract Personalized federated learning (pFL) collaboratively trains personalized models, which provides a customized model solution for individual clients in the presence of het- erogeneous distributed local data. Although many recent studies have applied various algorithms to enhance per- sonalization in pFL, they mainly focus on improving the performance from averaging or top perspective. How- ever, part of the clients may fall into poor performance and are not clearly discussed. Therefore, how to prevent these poor clients should be considered critically. Intu- itively, these poor clients may come from biased univer- sal information shared with others. To address this issue, we propose a novel pFL strategy, called Personalize Lo- cally, Generalize Universally (PLGU). PLGU generalizes the fine-grained universal information and moderates its bi- ased performance by designing a Layer-Wised Sharpness Aware Minimization (LWSAM) algorithm while keeping the personalization local. Specifically, we embed our proposed PLGU strategy into two pFL schemes concluded in this pa- per: with/without a global model, and present the training procedures in detail. Through in-depth study, we show that the proposed PLGU strategy achieves competitive general- ization bounds on both considered pFL schemes. Our exten- sive experimental results show that all the proposed PLGU based-algorithms achieve state-of-the-art performance.

1. Introduction Federated Learning (FL) is a popular collaborative re- search paradigm that trains an aggregated global learning model with distributed private datasets on multiple clients [16, 29]. This setting has achieved great accomplishments when the local data cannot be shared due to privacy and communication constraints [36]. However, because of the *Corresponding author.ExE 62 64 66 68 70 72 740.00.10.20.30.4 Poor Clients Poor ClientspFedMe FedRep Figure 1. Toy example in a heterogeneous pFL on CIFAR10, which includes 100 clients and each client obtains 3 labels. non-IID/heterogeneous datasets, learning a single global model to fit the “averaged distribution” may be difficult to propose a well-generalized solution to the individual client and slow the convergence results [24]. To address this problem, personalized federated learning (pFL) is devel- oped to provide a customized local model solution for each client based on its statistical features in the private train- ing dataset [5, 9, 11, 34]. Generally, we can divide existing pFL algorithms into two schemes: (I) with a global model [5, 23, 25, 40] or (II) without a global model [27, 28, 37]. Though many pFL algorithms make accomplishments by modifying the universal learning process [41, 47] or en- hancing the personalization [3, 27, 37], they may lead part of clients to fall into poor learning performance, where the personalization of local clients performs a large statistical deviation from the “averaged distribution”. To the best of our knowledge, none of the existing studies explore how to prevent clients from falling into poor personalized per- formance on these two schemes. For example, the poor medical learning models of some clients may incur seri- ous medical malpractice. To better present our concerned problem, we introduce a toy example in Figure 1, which is learned by two pFL algorithms representing these two schemes: pFedMe [40] and FedRep [3]. Though both al- gorithms achieve high averaged local model performance of 66.43% and 71.35%, there also 15% of clients are less than 64% and 14% clients are less than 69%, respectively. 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. 12167 This motivates us to exploit an effective strategy to prevent clients from falling into poor performance while without de- grading others, e.g., the green curve. Intuitively, we consider this phenomenon oftentimes comes from the biased universal information towards the clients with better learning performance. For scheme I, a simple-averaged aggregation may not perfectly handle data heterogeneity, as it generates serious bias between the global and local clients. For scheme II, abandoning the universal contribution may dismiss some information from other clients. Instead of designing a new pFL algorithm, we propose a novel pFL strategy on existing pFL stud- ies: generalizing the universal learning for unbiased local adaptation as well as keeping the local personalized fea- tures, called Personalize Locally, Generalize Universally (PLGU). The main challenge of PLGU is to generalize uni- versal information without local feature perturbation, as the statistical information is only stored locally. In this pa- per, we tackle this challenge by developing a fine-grained perturbation method called Layer-Wised Sharpness-Aware- Minimization (LWSAM) based on the SAM optimizer [7, 33], which develops a generalized training paradigm by leveraging linear approximation. Furthermore, we present how to embed this PLGU strategy with the perturbed uni- versal generalization on both the two pFL schemes. For scheme I (with the global model), we propose the PLGU-Layer Freezing (LF) algorithm. As illustrated in [21, 31, 48], each layer in a personalized model shares a different contribution: the shallow layers focus more on lo- cal feature extraction (personalization), and the deeper lay- ers are for extracting global features (universal). Specifi- cally, the PLGU-LF first explores the personalization score of each layer. Then, PLGU-LF freezes the important layer locally for personalization and uses the LWSAM optimizer with the consideration of obtained layer importance score for universal generalization. For scheme II (without the global model), we mainly focus on our proposed PLGU strategy FedRep algorithm [3], named PLGU-GRep. It gen- eralizes the universal information by smoothing personal- ization in the representation part. To show the extensibil- ity, we present that we can successfully extend our PLGU strategy to pFedHN [37], called PLGU-GHN, to improve learning performance, especially for poor clients. Further- more, we analyze the generalization bound on PLGU-LF, PLGU-GRep, and PLGU-GHN algorithms in-depth. Exten- sive experimental results also show that all three algorithms successfully prevent poor clients and outperform the aver- age learning performance while incrementally reducing the top-performance clients.

Li_SIM_Semantic-Aware_Instance_Mask_Generation_for_Box-Supervised_Instance_Segmentation_CVPR_2023

Abstract Weakly supervised instance segmentation using only bounding box annotations has recently attracted much re- search attention. Most of the current efforts leverage low- level image features as extra supervision without explicitly exploiting the high-level semantic information of the ob- jects, which will become ineffective when the foreground ob- jects have similar appearances to the background or other objects nearby. We propose a new box-supervised instance segmentation approach by developing a Semantic-aware In- stance Mask (SIM) generation paradigm. Instead of heav- ily relying on local pair-wise affinities among neighboring pixels, we construct a group of category-wise feature cen- troids as prototypes to identify foreground objects and as- sign them semantic-level pseudo labels. Considering that the semantic-aware prototypes cannot distinguish differ- ent instances of the same semantics, we propose a self- correction mechanism to rectify the falsely activated regions while enhancing the correct ones. Furthermore, to handle the occlusions between objects, we tailor the Copy-Paste operation for the weakly-supervised instance segmentation task to augment challenging training data. Extensive exper- imental results demonstrate the superiority of our proposed SIM approach over other state-of-the-art methods. The source code: https://github.com/lslrh/SIM .

1. Introduction Instance segmentation is among the fundamental tasks of computer vision, with many applications in autonomous driving, image editing, human-computer interaction, etc. The performance of instance segmentation has been im- proved significantly along with the advances in deep learn- ing [6, 12, 34, 38]. However, training robust segmentation networks requires a large number of data with pixel-wise annotations, which consumes intensive human labor and *denotes the equal contribution, †denotes the corresponding author. This work is supported by the Hong Kong RGC RIF grant (R5001-18). Figure 1. The pipeline of Semantic-aware Instance Mask (SIM) generation method. (a) shows the mask prediction produced by us- ing only low-level affinity supervision, where the foreground heav- ily blends with background. (b) and (c) show the semantic-aware masks obtained with our constructed prototypes, which perceive the entity of objects but are unable to separate different instances of the same semantics. (d) shows the final instance pseudo mask rectified by our proposed self-correction module. resources. To reduce the reliance on dense annotations, weakly-supervised instance segmentation based on cheap supervisions, such as bounding boxes [14,21,36], points [8] and image-level labels [1,18], has recently attracted increas- ing research attention. In this paper, we focus on box-supervised instance seg- mentation (BSIS), where the bounding boxes provide coarse supervised information for pixel-wise prediction task. To provide pixel-wise supervision, conventional methods [10, 19] usually leverage off-the-shelf proposal techniques, such as MCG [30] and GrabCut [31], to create pseudo instance masks. However, the training pipelines of these meth- ods with multiple iterative steps are cumbersome. Sev- eral recent works [14, 36] enable end-to-end training by taking pairwise affinities among pixels as extra supervi- sion. Though these methods have achieved promising per- formance, they heavily depend on low-level image features, such as color pairs [36], and simply assume that the proxi- mal pixels with similar colors are likely to have the same label. This leads to confusion when foreground objects have similar appearances to the background or other ob- 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. 7193 jects nearby, as shown in Fig. 1 (a). It is thus error-prone to use only low-level image cues for supervision since they are weak to represent the inherent structure of objects. Motivated by the fact that high-level semantic informa- tion can reveal intrinsic properties of object instances and hence provide effective supervision for segmentation model training, we propose a novel Semantic-aware Instance Mask generation method, namely SIM, to explicitly exploit the semantic information of objects. To distinguish proximal pixels with similar color but different semantics (please re- fer to Fig. 1 (a)), we construct a group of representative dataset-level prototypes, i.e., the feature centroids of differ- ent classes, to perform foreground/background segmenta- tion, producing semantic-aware pseudo masks (see Fig. 1 (b)). These prototypes abstracted from massive training data can capture the structural information of objects, en- abling more comprehensive semantic pattern understand- ing, which is complementary to affinity supervision of pair- wise neighboring pixels. However, as shown in Fig. 1 (c), these prototypes are unable to separate the instances of the same semantics, especially for overlapping objects. We consequently develop a self-correction mechanism to rec- tify the false positives while enhancing the confidence of true-positive foreground objects, resulting in more precise instance-aware pseudo masks, as shown in Fig. 1 (d). It is worth mentioning that our generated pseudo masks could co-evolve with the segmentation model without cum- bersome iterative training procedures in previous meth- ods [10, 21]. In addition, considering that the exist- ing weakly-supervised instance segmentation methods only provide very limited supervision for rare categories and overlapping objects due to the lack of ground truth masks, we propose an online weakly-supervised Copy-Paste ap- proach to create a combinatorial number of augmented training samples. Overall, the major contributions of this work can be summarized as follows: A novel BSIS framework is presented by developing a semantic-aware instance mask generation mechanism. Specifically, we construct a group of representative proto- types to explore the intrinsic properties of object instances and identify complete entities, which produces more reli- able supervision than low-level features. A self-correction module is designed to rectify the semantic-aware pseudo masks to be instance-aware. The falsely activated regions will be reduced, and the correct ones will be boosted, enabling more stable training and progressively improving the segmentation results. We tailor the Copy-Paste operation for weakly-supervised segmentation tasks in order to create more occlusion pat- terns and more challenging training data. The overall framework can be trained in an end-to-end manner. Ex- tensive experiments demonstrate the superiority of our method over other state-of-the-art methods.

Li_Regularize_Implicit_Neural_Representation_by_Itself_CVPR_2023

Abstract This paper proposes a regularizer called Implicit Neural Representation Regularizer (INRR) to improve the general- ization ability of the Implicit Neural Representation (INR). The INR is a fully connected network that can represent sig- nals with details not restricted by grid resolution. However, its generalization ability could be improved, especially with non-uniformly sampled data. The proposed INRR is based on learned Dirichlet Energy (DE) that measures similarities between rows/columns of the matrix. The smoothness of the Laplacian matrix is further integrated by parameterizing DE with a tiny INR. INRR improves the generalization of INR in signal representation by perfectly integrating the sig- nal’s self-similarity with the smoothness of the Laplacian matrix. Through well-designed numerical experiments, the paper also reveals a series of properties derived from INRR, including momentum methods like convergence trajectory and multi-scale similarity. Moreover, the proposed method could improve the performance of other signal representa- tion methods.

1. Introduction INR uses a fully connected network (FCN) ϕθ(x) : Rd7→Roto approximate the explicit solution of an implicit function F x, ϕθ,∇xϕθ,∇2 xϕθ, . . .
= 0. For an example, we can represent a gray-scale image X∈Rm×nwith an INRϕθ(x) :R27→Rwhich satisfied ϕθ(i m,j n) =Xij, i∈ {1, . . . , m }, j∈ {1, . . . , n }. Compared with traditional grid representation X, INR’s representation ability to details is not restricted by grid resolution m, n as INR can predict the pixel value at any location (x, y)∈R2even not equals to (i m,j n). Besides the representation ability of INR, generalization ability is critical for a neural network. We explore the em- *This work was supported by the National Key Research ,Development Program (2020YFA0713504), the National Natural Science Foundation of China (61977065) and the Macao Science and Technology Development Fund (061/2020/A2).pirical generalization ability via a 256×256 gray-scale non-uniformly sampled image inpainting task as Figure 2(a) shows. Although INR fits training data perfectly in Fig- ure 2(b), its prediction outside training data is unreasonable. Theoretical analysis of INR illustrates that a hyper-parameter controls the smoothness degree of ϕθ(x). Moreover, the ex- periments show that the best hyper-parameter varies with the missing rate (the percentage of unsampled pixels) as Fig- ure 3 shows. Adjusting this hyper-parameter cannot make the non-uniformly missing case perform best, as different locations might have different missing rates. A carefully designed regularizer is proposed to improve the generalization ability of INR. It is based on Adaptive and Implicit Regularization (AIR) which is a learned Dirichlet Energy (DE) [12] that measures similarities or correlations between rows/columns of X. The smoothness of the Lapla- cian matrix is further integrated by parameterizing DE with a tiny INR. The structure of the proposed implicit neural representation regularizer (INRR) is shown in Figure 1(b). Because a smooth Laplacian matrix represents non-local prior and large-scale local prior in vision data, INRR can improve the generalization of INR in image representation. Numerous numerical experiments show that INRR outper- forms various classical regularizers, including total variation (TV), L2energy, and so on. As a regularizer both in a new form and with new meaning, INRR can be combined with other signal representation methods, such as deep matrix factorization (DMF) [1]. To summarize, the contributions of our work include the following: •Neural Tangent Kernel (NTK) [1] theoretically analyzes the generalization ability of INR and why INR performs poorly with nonuniform sampling is given. •A tiny INR parameterized regularizer named INRR is proposed based on DE, which perfectly integrates the image’s self-similarity with the smoothness of the Laplacian matrix. •A series of properties derived from INRR, including momentum methods, multi-scale similarity, and 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. 10280 𝑥𝑥 𝑦𝑦 𝑓𝑓(𝑥𝑥,𝑦𝑦) 𝑥𝑥 𝑔𝑔(𝑥𝑥)(a) (b)INR INRR(c) 𝑓𝑓(𝑥𝑥,𝑦𝑦)𝑥𝑥𝑦𝑦 𝐿𝐿(𝑔𝑔𝑇𝑇𝑦𝑦1𝑔𝑔(𝑦𝑦2))𝑦𝑦1𝑦𝑦2 INR-Z𝑥𝑥𝑦𝑦 𝑓𝑓(𝑁𝑁𝑥𝑥,𝑦𝑦) ℎ𝑥𝑥,𝑦𝑦,𝑓𝑓𝑁𝑁𝑥𝑥,𝑦𝑦0.00 -0.02 -0.04 -0.06 -0.08 -0.10Figure 1. Overview of proposed improve scheme for INR. (a) INR is a fully connected neural network which maps from coordinate to pixel value. (b) INRR is a regularization term represented by an INR which can capture the self-similarity. (c) INR-Z improve the performance of INR by combining the neighbor pixels with coordinate together as the input of another INR. (a) Sampling (b) INR (18.1 dB) (c) INRR (23.3 dB) Figure 2. Image fitting results. All the methods are based on the SIREN to fit an 256×256Baboon with the sampling data in (a). (b) trained with a vanilla SIREN while (c) trained with proposed INRR. alization ability, are revealed by well-designed numeri- cal experiments.

Long_CapDet_Unifying_Dense_Captioning_and_Open-World_Detection_Pretraining_CVPR_2023

Abstract Benefiting from large-scale vision-language pre-training on image-text pairs, open-world detection methods have shown superior generalization ability under the zero-shot or few-shot detection settings. However, a pre-defined cate- gory space is still required during the inference stage of ex- isting methods and only the objects belonging to that space will be predicted. To introduce a “real” open-world de- tector, in this paper, we propose a novel method named CapDet to either predict under a given category list or di- rectly generate the category of predicted bounding boxes. Specifically, we unify the open-world detection and dense caption tasks into a single yet effective framework by in- troducing an additional dense captioning head to gener- ate the region-grounded captions. Besides, adding the cap- tioning task will in turn benefit the generalization of detec- tion performance since the captioning dataset covers more concepts. Experiment results show that by unifying the dense caption task, our CapDet has obtained significant performance improvements (e.g., +2.1% mAP on LVIS rare classes) over the baseline method on LVIS (1203 classes). Besides, our CapDet also achieves state-of-the-art perfor- mance on dense captioning tasks, e.g., 15.44% mAP on VG V1.2 and 13.98% on the VG-COCO dataset.

1. Introduction Most state-of-the-art object detection methods [ 33,34, 50] benefit from a large number of densely annotated detec- tion datasets ( e.g., COCO [ 27], Object365 [ 36], LVIS [ 12]). However, this closed-world setting results in the model only being able to predict categories that appear in the training set. Considering the ubiquity of new concepts in real-world scenes, it is very challenging to locate and identify these new visual concepts. This predictive ability of new concepts in open-world scenarios has very important research value *Equal contribution. †Corresponding authors. sofa tablebottlesuitcase (b) OVD (a) OWD dense captioning head alignment a white wall outlet a black lamp bottle region embeddings remote, table, bottle, suitcase, sofa,…pre-defined category list (C)CapDet (ours) unknown unknownremote lamp text encoderFigure 1. Comparison of the different model predictions under OWD, OVD, and our setting. (a) OWD methods [ 14,18,48] are not able to describe the detailed category of the detected unknown objects and (b) the performance of OVD methods [ 8,12,41] usu- ally depends on the pre-defined category list during the inference. (c) With the unification of two pipelines of dense captioning and open-world detection pre-training, our CapDet can either predict under a given category list or directly generate the description of predicted bounding boxes. in real-world applications such as object search [ 29,30], in- stance registration [ 45], and human-object interaction mod- eling [ 10]. Currently, the open world scenario mainly includes two tasks: open world object detection [18] (OWD) and open- vocabulary object detection [44] (OVD). Although the paradigms of OWD and OVD tasks are closer to the real world, the former cannot describe the specific concept of the detected unknown objects and requires a pre-defined category list during the inference. Specifically, 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. 15233 in Figure 1, previous OWD methods [ 14,18,48] would rec- ognize new concepts not in the predefined category space as “unknown”. Further, another line of task OVD requires the model to learn a limited base class and generalize to novel classes. Compared to the zero-shot object detection (ZSD) proposed by [ 32], OVD allows the model to use ex- ternal knowledge, e.g., knowledge distillation from a large- scale vision-language pre-trained model [ 8,12], image- caption pairs [ 44], image classification data [ 49], grounding data [ 25,41,46]. With the external knowledge, OVD meth- ods show a superior generalization capacity to detect the novel classes within a given category space. However, as shown in Figure 1, when given an incomplete category list, OVD can only predict the concepts that appear in the given category list, otherwise, there will be recognition errors, ( i.e., as illustrated in Figure 1(b), the OVD methods prone to predict the “wall socket” as “remote”, since the latter is in the category list but not the former). Thus, under the OVD setting, we mainly face the follow- ing two challenges: ( i) it is difficult to define a complete list of categories; ( ii) low response values on rare categories often lead to recognition errors. This is mainly because we cannot exhaustively enumerate new objects in the real world, and secondly, it is difficult to collect enough sam- ples for rare classes. However, the fact that rare objects in the real world, even some new objects that are unknown to humans, such as UFOs, do not prevent people from using natural language to describe it as “a flying vehicle that looks like a Frisbee”. Therefore, based on the above observations, in this pa- per, we consider a new setting that is closer to the open world and real scenes, i.e., we expect the model to both detect and recognize concepts in a given category list, and to generate corresponding natural language descriptions for new concepts or rare categories of objects. Early dense cap- tioning methods [ 9,17] can locate salient regions in images and generate the region-grounded captions with natural lan- guage. Inspired by this, to address the challenges faced in the OVD setting, we propose to unify the two pipelines of dense captioning and open-world detection pre-training into one training framework, called CapDet . It empowers the model with the ability to both accurately detect and recog- nize common object categories and generate dense captions for unknown and rare categories by unifying the two train- ing tasks. Specifically, our CapDet constructs a unified data for- mat for the dense captioning data and detection data. With the data unification, CapDet further adopts a unified pre- training paradigm including open-world object detection and dense captioning pre-training. For open-world detec- tion pretraining, we treat the detection task as a semantic alignment task and adopt a dual encoder structure as [ 41] to locate and predict the given concepts list. The conceptslist contains category names in detection data and region- grounded captions in dense captioning data. For dense cap- tioning pretraining, CapDet proposes a dense captioning head to take the predicted proposals as input to generate the region-grounded captions. Due to the rich visual con- cepts in the dense captioning data , the integration of dense captioning tasks will in turn benefit the generalization of detection performance. Our experiments show that the integration of few dense captioning data brings in large improvement in the object detection datasets LVIS, e.g., +2.7% mAP on LVIS. The unification of dense captioning and detection pre-training gains an additional 2.3% increment on LVIS and 2.1% in- crement on LVIS rare classes. Besides, our model also achieves state-of-the-art performance on dense captioning tasks. Note that our method is the first to unify dense cap- tioning and open-world detection pretraining. To summarize, our contributions are three folds: • We propose a novel open-vocabulary object detection framework CapDet, which cannot only detect and rec- ognize concepts in a given category list but also gen- erate corresponding natural language descriptions for new concept objects. • We propose to unify the two pipelines of dense cap- tioning and open-world detection pre-training into one training framework. Both two pre-training tasks are beneficial to each other. • Experiments show that by unified dense captioning task and detection task, our CapDet gains significant performance improvements on the open-vocabulary object detection task ( e.g., +3.3% mAP on LVIS rare classes). Furthermore, our CapDet also achieves state- of-the-art performance on the dense captioning tasks, e.g., 15.44% mAP on Visual Genome (VG) V1.2 and 13.98% mAP on VG-COCO.

Narayan_DF-Platter_Multi-Face_Heterogeneous_Deepfake_Dataset_CVPR_2023

Abstract Deepfake detection is gaining significant importance in the research community. While most of the research ef- forts are focused towards high-quality images and videos with controlled appearance of individuals, deepfake gener- ation algorithms now have the capability to generate deep- fakes with low-resolution, occlusion, and manipulation of multiple subjects. In this research, we emulate the real- world scenario of deepfake generation and propose the DF- Platter dataset, which contains (i) both low-resolution and high-resolution deepfakes generated using multiple genera- tion techniques and (ii) single-subject and multiple-subject deepfakes, with face images of Indian ethnicity. Faces in the dataset are annotated for various attributes such as gen- der, age, skin tone, and occlusion. The dataset is prepared in 116 days with continuous usage of 32 GPUs account- ing to 1,800 GB cumulative memory. With over 500 GBs in size, the dataset contains a total of 133,260 videos en- compassing three sets. To the best of our knowledge, this is one of the largest datasets containing vast variability and multiple challenges. We also provide benchmark re- sults under multiple evaluation settings using popular and state-of-the-art deepfake detection models, for c0 images and videos along with c23 and c40 compression variants. The results demonstrate a significant performance reduc- tion in the deepfake detection task on low-resolution deep- fakes. Furthermore, existing techniques yield declined de- tection accuracy on multiple-subject deepfakes. It is our assertion that this database will improve the state-of-the- art by extending the capabilities of deepfake detection al- gorithms to real-world scenarios. The database is available at: http://iab-rubric.org/df-platter-database.

1. Introduction With the advent of diverse deep learning architectures, significant breakthrough have been made in the field of im- age/video forgery. This has led to an incredible rise in the *Equal contribution by student authors. (c) (a) (b) Figure 1. Samples showcasing multi-face deepfakes circulated on social media. (a) A zoom call with a deepfake of Elon Musk [8] (b) Real-time deepfake generation at America’s Got Talent [9] (c) Deepfake round-table with multiple deepfake subjects [33]. amount of fake multimedia content being generated due to increased accessibility and less training requirements. Not only has the amount of such media risen, but the sophisti- cation of such content has also improved drastically, mak- ing it indistinguishable from real videos. While most deep- fakes are used for entertainment purposes like parody films and filters in apps, they can also be used to illicitly defame someone, spread misinformation or propaganda, or conduct fraud. In 2020 Delhi state elections in India, a deepfake video of a popular political figure was created [34] and ac- cording to some estimates, the deepfake was disseminated to about 15 million people in the state [13]. Given the abuse of deepfakes and their possible impact, the necessity for bet- ter and robust deepfake detection methods is unavoidable. Designing a dependable deepfake detection system re- quires availability of comprehensive deepfake datasets for training. Table 1 summarizes the key characteristics of the publicly available deepfake datasets. Most of the datasets 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. 9739 Table 1. Quantitative comparison of DF-Platter with existing Deepfake datasets. DatasetReal VideosFake VideosTotal VideosTotal SubjectsReal SourceMultiple faces per image/videoFace OcclusionGeneration TechniquesLow Resolution1Annotations2 FF++ [30] 1,000 4,000 5,000 N/A YouTube ✗ ✗ 4 ✗ ✗ Celeb-DF [21] 590 5,639 6,229 59 YouTube ✗ ✗ 1 ✗ ✗ UADFV [36] 49 49 98 49 YouTube ✗ ✗ 1 ✗ ✗ DFDC [6] 23,654 104,500 128,154 960 Self-Recording ✗ ✗ 8 ✗ ✗ DeepfakeTIMIT [15] 640 320 960 32 VidTIMIT ✗ ✗ 2 ✓ ✗ DF-W [29] N/A 1,869 1,869 N/A YouTube & Bilibili ✗ ✗ 4 ✗ ✗ KoDF [16] 62,166 175,776 237,942 403 Self-Recording ✗ ✗ 6 ✗ ✗ WildDeepfake [39] 707 707 1,414 N/A Internet ✗ ✗ N/A ✗ ✗ OpenForensics [17] 45,473* 70,325* 115,325* N/A Google Open Images ✓ ✓ 1 ✗ ✗ DeePhy [26] 100 5,040 5,140 N/A YouTube ✗ ✓ 3 ✗ ✓ DF-Platter (ours) 764 132,496 133,260 454 YouTube ✓ ✓ 3 ✓ ✓ 1Low resolution means the dataset contains low-resolution deepfakes generated using low-resolution videos and not by down-sampling. 2The dataset provides annotations such as skin tone, facial attributes and face occlusion. *The number of images have been reported since the dataset contains only images. contain high-resolution images with s