FraunhoferIOSB/Synset-Boulevard

Dataset Card for Synset Boulevard

The Synset Boulevard dataset is a synthetic dataset for the task of vehicle make and model recognition (VMMR), generated using path tracing and physically-based, data-driven models, and contains 259,200 images with 162 different vehicle models of 43 makes depicted in front view.

Download / more info: synset.de/datasets/synset-blvd/

Dataset Details

Dataset Description

The Synset Boulevard dataset is designed for the task of vehicle make and model recognition (VMMR), and is—to the best of our knowledge—the first entirely synthetically generated large-scale VMMR image dataset.

Through the simulation of image data rather than the manual annotation of real data, it is intended to mitigate common challenges in state-of-the-art VMMR datasets, namely bias, human error, privacy, and the challenge of providing systematic updates. On the other hand, the provision and use of synthetic data introduce individual challenges, such as potential domain gaps and a less pronounced intra-class variance.

The dataset was generated using path tracing and physically-based, data-driven models, and contains 32,400 independent images (each with different imaging simulations and with/without masked license plates, leading to a total of 259,200 images) from 162 different vehicle models of 43 makes depicted in front view. It is split into 8 sub-datasets to investigate the influence of optical/imaging effects on the classification ability.

• Curated by: Anne Sielemann, Stefan Wolf, Jens Ziehn, Masoud Roschani, and Juergen Beyerer. Fraunhofer IOSB, Germany.

• Funded by:

  • Fraunhofer Internal Programs under Grant No. PREPARE 40-02702 within the ML4Safety project
  • the Ministry of Economic Affairs, Labour and Housing of the state of Baden-Wuerttemberg, Germany, as part of the FeinSyn research project

• License: CC-BY 4.0

Dataset Sources

• Repository: synset.de/datasets/synset-blvd/
• Paper: Sielemann, A., Wolf, S., Roschani, M., Ziehn, J. and Beyerer, J. (2024). Synset Boulevard: A Synthetic Image Dataset for VMMR. In 2024 IEEE International Conference on Robotics and Automation (ICRA).

Uses

The dataset is designed for the task of vehicle make and model recognition (VMMR), containing surveillance-type images of vehicles in front view, corresponding to the perspective of traffic cameras.

Direct Use

The dataset is intended for the following use cases:

• Training ML models for the task of VMMR
• Analyzing the difference between the synthetic dataset and real-world VMMR datasets, especially the closely related CompCars Surveillance dataset
• Testing ML models for the task of VMMR, in particular by using additional environmental information per image

Out-of-Scope Use

The dataset should not be used for critical applications, particularly high-risk applications as named by the European AI Act under Annex III (which includes "AI systems intended to be used for the ‘real-time’ and ‘post’ remote biometric identification of natural persons" and "AI systems intended to be used as safety components in the management and operation of road traffic"), without exhaustive research into the fitness of the dataset, to evaluate whether it is "relevant, sufficiently representative, and to the best extent possible free of errors and complete in view of the intended purpose of the system." No such claim is not made with the publication of this dataset.

Dataset Structure

The dataset is separated into the following variants:

• Variants (concerning the presence of license plates):

  • Original: Contains raw images including license plates as shipped with the original 3D model. License plate appearance is identical per vehicle class
  • MaskedLicensePlates (MLP): Raw images with license plate regions detected a-posteriori via ML, and replaced by the average image color (a shade of gray)

• Qualities (concerning imaging artifacts):

  • Good: Contains images with well-chosen automatic exposure control (AEC) and white balance (AWB) values and low noise levels, corresponding to a well-calibrated camera
  • Bad: Contains images with stochastically deviating AEC and AWB, higher noise levels and lens flares

• Bayer pattern artifacts:

  • Bayer: Contains artifacts from Bayer pattern demosaicing
  • Regular: Contains no Bayer pattern artifacts

Each of these 8 variants uses the same 32,400 geometric path tracing renderings in high dynamic range as a basis for the simulation of imaging artifacts. Their combination leads to 259,200 raw images total.

The 32,400 geometric renderings are subdivided into 162 models over 43 makes.

Each of the 32,400 renderings contains the raw image (i.e., the simulated camera image), a label image for semantic segmentation, and metadata about the environment (environment.csv) and car paint colors (vehicle-colors.csv).

Each of the 162 classes contains data about the vehicle from the ADAC database (General German Automobile Club), including official model names, model years and number of doors.

The dataset provides an exemplary train and test split (3:1) given by train.csv and test.csv.

Dataset Creation

Curation Rationale

Through the simulation of image data rather than the manual annotation of real data, it is intended to mitigate common challenges in state-of-the-art VMMR datasets, namely bias, human error, privacy, and the challenge of providing systematic updates. On the other hand, the provision and use of synthetic data introduce individual challenges, such as potential domain gaps and a less pronounced intra-class variance.

The dataset was designed to be comparable to the CompCars Surveillance dataset, as the currently largest known public dataset for VMMR for frontal traffic camera perspectives.

The goal of the dataset was in particular to describe all known properties of the dataset as far as possible through transparent stochastic processes, enabling an understanding of the generation / synthesis principles and allowing to evaluate ML models based on detailed properties of the data.

Source Data

• The dataset was generated in the OCTANE simulation framework, particularly using path tracing through the Blender Cycles engine.

• The 3D models of the vehicles were collected from commercial sources, mainly Dosch Design.

• Vehicle paint colors were sampled stochasically based on a report by the German Federal Motor Transport Authority (KBA) showing new vehicle registrations by color in Germany (2021), based on 2,622,132 newly registered cars. For full details see the website or publication.

• Image-based lighting (IBL) uses 183 environment maps from PolyHaven

• Road surface textures comprise 150 different textures with four different types of markings (dashed vs. solid, white vs. yellow). Textures were generated stochastically based on texturelib.com

Data Collection and Processing

[More Information Needed]

Who are the source data producers?

• Dosch Design as the producer of the majority of 3D models is a company in Marktheidenfeld, Germany, focusing on 3D models for visualization and design.

• PolyHaven as the provider of the environment maps for image-based lighting (IBL) is an online library for open (CC0) 3D assets provided by different authors.

• texturelib.com as the provider for the open base road textures is operated by Dmitriy Chugai and Oleksandr Chugai.

Annotations

Annotation process

The major part of annotations, including vehicle colors, segmentation masks and environmental conditions is based on ground truth data created as part of the scene generation / rendering process. Semantic segmentation images were rendered using the Ogre 3D rendering engine plugin to OCTANE, which provides rasterization / shading-based image generation.

The only manual annotation performed in the creation of the particular dataset is the mapping between 3D models and the corresponding vehicle make and model information (performed once for each class). This was performed using a comparison of the rendered images with real images from the database of the ADAC (General German Automobile Club).

Who are the annotators?

The annotation of 3D models to vehicle make and model information was performed by the authors.

Personal and Sensitive Information

The dataset contains no data that might be considered personal, sensitive, or private.

Bias, Risks, and Limitations

• Model variants: Every class in the dataset contains only a single variant, while in practice, optional or retrofitted equipment such as sunroofs or fog lights may vary within one make/model/year class. The accuracy of the 3D models is not evaluated separately. Different national variants of the vehicles are not reflected in the dataset. The dataset covers European, Asian and American models, but has a strong focus on models and variants common in Western Europe.

• Vehicle lights: The 3D models were not annotated for individual vehicle light functions, such that no distinction between daytime lights, high beam, turn indicators, etc., is made in the dataset. The lights are not triggered individually.

• License plates: License plates are modeled as part of the 3D mesh and textures and are therefore fixed for each vehicle geometry. Some vehicles feature fixed license plate numbers or logos, others contain empty license plates or no license plates. Therefore, in the dataset without masked license plates, license plate appearance will be identical across cars of the same class, and shared among some different classes as well.

• Environment: Environment variation is limited to over- all lighting conditions and road model and textures. No com- plex shadows or reflections from roadside objects, other ve- hicles, occlusions, environment conditions (snow, raindrops, fog, ...) or low light / nighttime conditions are included.

• Environment and lighting: Available light models are currently limited and not calibrated. Therefore, no absolute scales are given and the relative vehicle light brightness (and corresponding effects) will be incorrect. Surface properties for physically-based rendering are selected qualitatively and are not based on accurate physical measurements.

• Perspective and camera: Only frontal perspective images are included in the dataset, and only one set of intrinsic camera parameters is used, and only a single camera lens type (based on a Tamron M112FM35 35 mm lens) and only a very limited set of imaging artifacts are simulated.

Recommendations

It is recommended to use the dataset primarily for scientific research. Application to practical real-world use cases should include human oversight and the exhaustive evaluation of the fitness for the respective purpose, including the impact of domain shifts.

Citation

BibTeX:

@inproceedings{synset_blvd_sielemann_2024,
  title={{Synset Boulevard: A Synthetic Image Dataset for VMMR}},
  author={Sielemann, Anne and Wolf, Stefan and Roschani, Masoud and Ziehn, Jens and Beyerer, Juergen},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  year={2024}
}

APA:

Sielemann, A., Wolf, S., Roschani, M., Ziehn, J. & Beyerer, J. (2024). Synset Boulevard: A Synthetic Image Dataset for VMMR. 2024 IEEE International Conference on Robotics and Automation (ICRA).

Dataset Card Authors

Anne Sielemann, Stefan Wolf, Jens Ziehn, Masoud Roschani, and Juergen Beyerer

Dataset Card Contact

Anne Sielemann
Fraunhofer IOSB
Group »Automotive and Simulation«
Fraunhoferstr. | 76131 Karlsruhe | Germany
anne.sielemann@iosb.fraunhofer.de
www.iosb.fraunhofer.de

Jens Ziehn
Fraunhofer IOSB
Group leader »Automotive and Simulation«
Fraunhoferstr. | 76131 Karlsruhe | Germany
Phone +49 721 6091 – 633
jens.ziehn@iosb.fraunhofer.de
www.iosb.fraunhofer.de