FraunhoferIOSB/Synset-Background-Effect-Datasets

Synset Background Effect Datasets

For investigating the effect of background on feature importance and classification performance, we systematically generated six synthetic datasets for the task of traffic sign recognition, which differ only in their degree of camera variation and background correlation. Each of these datasets contains 82 classes of traffic signs with 1,100 images per class, resulting in 90,200 images per dataset, summing up to a total of 541,200 images.

Website: synset.de/datasets/synset-signset-ger/background-effect/
Paper: Sielemann, A., Barner, V., Wolf, S., Roschani, M., Ziehn, J. and Beyerer, J. (2025). Measuring the Effect of Background on Classification and Feature Importance in Deep Learning for AV Perception. In 2025 IEEE International Automated Vehicle Validation Conference (IAVVC). [arXiv]
Authors: Anne Sielemann, Valentin Barner, Stefan Wolf, Masoud Roschani, Jens Ziehn, and Juergen Beyerer. Fraunhofer IOSB, Germany.
Funded by: Fraunhofer Internal Programs under Grant No. PREPARE 40-02702 within the ML4Safety project and the German Federal Ministry for Economic Affairs and Climate Action, within the program “New Vehicle and System Technologies” as part of the AVEAS research project.
License: CC-BY 4.0

Description

Common approaches to explainable AI (XAI) for deep learning (DL)-based image classification focus on analyzing the importance of input features on the classification task in a given model: saliency methods like SHAP and GradCAM are used to measure the impact of spatial regions of the input image on the classification result. Combined with ground truth information about the location of the object in the input image, for example, a binary mask, it is determined whether object pixels had a high impact on the classification result, or whether the classification focused on background pixels. The former is considered to be a sign of a healthy classifier, whereas the latter is assumed to suggest overfitting on spurious correlations.

A major challenge, however, is that these intuitive interpretations are difficult to test quantitatively, and hence the output of such explanations lacks explanation itself. One particular reason is that correlations in real-world data are difficult to avoid, and whether they are spurious or legitimate is debatable. To shed light on this issue and test whether feature importance-based XAI reliably distinguishes between true learning and problematic overfitting, we utilize the task of traffic sign recognition and systematically generated six synthetic datasets, which only differ in their degree of camera variation and background correlation. Thereby, a correlated background means that each traffic sign is depicted in its most probable environment according to German traffic code / regulation StVO (Straßenverkehrs-Ordnung) categorized in "urban", "nature", and "urban and nature". A traffic sign warning of wildlife crossing is, for example, likely to be set up on a rural road with natural background, while a sign warning of pedestrians is probable to be placed in an urban context. An uncorrelated background, however, means that the background is randomly chosen and thus not semantically linked to the depicted traffic sign class.

For dataset generation, we utilized our parameterizable rendering pipeline from our work on the Synset Signset Germany dataset. The pipeline is based on the Fraunhofer simulation platform OCTAS. The dataset consists of six subdatasets: correlated and uncorrelated backgrounds cross the camera variation stages frontal, medium and high. Each of these datasets contains 82 classes of traffic signs with 1,100 images per class, resulting in 90,200 images per dataset, summing up to a total of 541,200 images. The images were rendered with the rasterization-based engine OGRE3D.

Citation

To cite this dataset in your scientific work, please use the following bibliography entry:

BibTeX:

@inproceedings{measuring_effect_of_background_sielemann_2025,
  title={{Measuring the Effect of Background on Classification and Feature Importance in Deep Learning for AV Perception}},
  author={Sielemann, Anne and Barner, Valentin and Wolf, Stefan and Roschani, Masoud and Ziehn, Jens and Beyerer, Juergen},
  booktitle={2025 IEEE International Automated Vehicle Validation Conference (IAVVC)},
  year={2025}
}

APA:

Sielemann, A., Barner, V., Wolf, S., Roschani, M., Ziehn, J., and Beyerer, J. (2025). <br>
Measuring the Effect of Background on Classification and Feature Importance in Deep Learning for AV Perception. <br>
In 2025 IEEE International Automated Vehicle Validation Conference (IAVVC).

In case of copying and redistributing, or publishing an adapted version of our dataset, please provide the name of our dataset, the creator names, a copyright notice, a link to this website, a license notice with a link to the license, and, if changes were made, a disclaimer notice, and a short description of the applied changes. For example, as follows:

This work is based on Measuring the Effect of Background on Classification and Feature Importance in Deep Learning for AV Perception
by Anne Sielemann, Valentin Barner, Stefan Wolf, Masoud Roschani, Jens Ziehn, and Juergen Beyerer,
© 2025 Fraunhofer IOSB, All rights reserved.
Link: https://synset.de/datasets/synset-signset-ger/background-effect/
Licence: CC BY 4.0, https://creativecommons.org/licenses/by/4.0/
Disclaimer: The original authors are neither affiliated nor responsible for any applied changes.

Uses

The dataset was designed for the investigation of the effect of background correlations on the classification performance and the spatial distribution of important classification features within the task of traffic sign recognition.

Direct Use

The dataset is intended for the following use cases:

• Training ML models for the task of German traffic sign recognition.
• Analyzing the difference between the synthetic dataset and real-world traffic sign recognition datasets, especially the closely related GTSRB dataset.
• Investigating the effects of background correlations.

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 Synset Background Effect Datasets include six dataset variants:

1. Correlated, frontal
2. Correlated, medium
3. Correlated, high
4. Uncorrelated, frontal
5. Uncorrelated, medium
6. Uncorrelated, high

Thereby, correlated means, that the used environment maps for image-based lighting are correlated to the depicted traffic sign class, while in the uncorrelated case the environment maps are randomly chosen for each image (independent from the traffic sign class). Frontal, medium, and high refer to the level of camera variation. For the frontal case, all traffic signs are soley pictured in frontal perspective without any camera variation. For the remaining two levels, the camera roll, pitch, and yaw angles are normally distributed with mean equal to zero. For medium, the standard deviations are set to 1.5° (roll), 5.0° (pitch), and 13.333° (yaw). In case of high, the standard deviations are increased to 3.0° (roll), 10.0° (pitch), and 26.666° (yaw).

Each of these datasets contains 82 classes of traffic signs with 1,100 images per class, resulting in 90,200 images per dataset, summing up to a total of 541,200 images. However, in addition to each of these raw images (i.e., the simulated camera image) we provide a semantic segmentation image, a mask image, and metadata about the traffic sign status (orientation, upper signs, lower signs, etc.), the environment (daytime, contrast, location, etc.), and the imaging effects (noise level, motion blur strength, aec error, etc.).

The datasets provide exemplary training and validation splits (500 training and 600 testing images per dataset and class).

Dataset Creation

Curation Rationale

The use case of traffic sign recognition has the advantages of, on the one hand, representing a well-understood and established task that provides a wide range of publicly available datasets and applicable models. On the other hand, it remains the subject of active research, in particular, to address challenges such as corner cases and weather conditions, and it has practical relevance, for example, for driver assistance systems, automated driving, and mapping. Since new traffic signs are constantly being released (2020 in Germany) and the coverage of existing signs in publicly available datasets is still limited for a distinction of less common classes, the demand for both training and testing data still persists.

The dataset was designed, to enable the investigation of background effects on the classification performance and neural network attention.

Source Data

• The dataset was generated in the OCTAS® simulation framework, by using rasterization trough the OGRE engine.

• The traffic sign template images, which are used as input to the GAN-based texture synthesis, stem from the Wikipedia overview of German traffic signs.

• Image-based lighting (IBL) uses 140 environment maps from PolyHaven.

• The 3D geometry of the tree that serves as an occlusion object originates from PolyHaven.

Who are the source data producers?

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

• Wikipedia, one of the largest free multilingual open-content encyclopedias, includes the complete list of existing German traffic signs and their template images.

Annotations

Annotation process

All annotations, including masks, segmentation images, camera parameters and artifacts, and environmental conditions, are based on ground truth data created as part of the scene generation / rendering process. Semantic segmentation images were rendered using the Ogre rendering engine plugin for OCTAS®, which provides rasterization / shading-based image generation. The environment labels stem from PolyHaven.

Personal and Sensitive Information

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

Bias, Risks, and Limitations

• Traffic Signs: The wear and tear generation is limited to artifacts such as color fading, scratches, screw holes, and sticker residues. Complex stickers, graffiti, or dirt are not included. Retroreflector patterns are excluded, and retroreflection is not simulated. The traffic signs are solely mounted on metallic traffic sign poles.

• Environment: Environmental variation includes no adverse weather conditions (snow, raindrops, fog, ...).

• Occlusions: All included occlusions or shadows stem from a single 3D tree geometry.

• Camera: Only one set of intrinsic camera parameters is used, and only a single camera lens type (based on a Tamron M112FM35 35 mm lens) is simulated. It can be assumed that the set of simulated imaging artifacts is not complete.

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.

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