The ATLAS of Traffic Lights

2025/02/10 by Rupert Polley, Polley, Rupert, Nikolai Polley +9 · 1 citation
Computer Science · Engineering · #Image Enhancement Techniques #Autonomous Vehicle Technology and Safety #Advanced Neural Network Applications

paper · pdf · doi:10.48550/arxiv.2504.19722

Abstract

Changelog: v1.0 - 29. Apr. 2025: The associated paper has been accepted for publication at IEEE IV 2025, and a preprint is available on arXiv. No changes were made to the dataset. v0.9 - 10. Feb. 2025: We released the dataset with over 33,000 labeled images. Until the paper associated with this dataset is published, we may modify its structure and update, add, or replace images and labels. v0.1 - 31. Jan. 2025: We released a small dataset sample. ATLAS ATLAS (Applied Traffic Light Annotation Set) is a new, publicly available dataset designed to improve traffic light detection for autonomous driving. Existing open-source datasets often omit certain traffic light states and lack camera configurations for near and far distances. To address this, ATLAS features over 33,000 images collected from three synchronized cameras—wide, medium, and tele—with varied fields of view in the German city of Karlsruhe. This setup captures traffic lights at diverse distances and angles, including difficult overhead views. Each of the dataset’s 72,998 bounding boxes is meticulously labeled for 25 unique pictogram-state classes, covering rare but critical states (e.g., red-yellow) and pictograms (straight-right, straight-left). Additional annotations include challenging conditions such as heavy rain. All data is anonymized using state-of-the-art tools. ATLAS provides a comprehensive, high-quality resource for robust traffic light detection, overcoming the limitations of existing datasets. The associated paper, Polley et al., "The ATLAS of Traffic Lights: A Reliable Perception Framework for Autonomous Driving", has been accepted for publication at IEEE IV 2025, and a preprint is available on ArXiv (arXiv:2504.19722). Camera FOV [°] Resolution Images Front-Medium 61 × 39 1920 × 1200 25,158 Front-Tele 31 × 20 1920 × 1200 5,109 Front-Wide 106 × 92 2592 × 2048 2,777 Directory Format: We provide the dataset in the following format: ├── ATLAS ├── train ├── front_medium ├── images ├── front_medium₁722622455-950002160.jpg ├── labels ├── front_medium₁722622455-950002160.txt ├── front_tele ├── front_wide ├── test ├── ATLAS_classes.yaml ├── LICENSE └── README.md We provide the dataset in split ZIP files (.zip, .z01, .z02, etc.). The standard unzip tool does not handle split archives well, we recommend 7-Zip for extraction (e.g.: 7z x ATLAS.zip). If 7-Zip is not available, first merge the files and then unzip them: zip -s 0 ATLAS.zip --out ATLAS_merged.zip && unzip ATLAS_merged.zip Annotation Format: Each line in an annotation file describes one bounding box using five fields: class_id x_center y_center width height class_id: An integer indicating the class of the annotated object. The file ATLAS_classes.yaml contains human-readable names corresponding to each numeric label. x_center, y_center: The normalized coordinates of the bounding box center, relative to the image dimensions (in the range [0,1]), where x_center is measured horizontally and y_center vertically. width, height: The normalized width and height of the bounding box are also expressed in the range [0,1]. These values are obtained by dividing the bounding box width and height in pixels by the overall image width and height, respectively. Terms and Conditions The ATLAS dataset provided by the FZI Research Center for Information Technology is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0). The dataset may be used exclusively for non-commercial purposes, including teaching and scientific research. The dataset is provided "as is." Although the Licensor has made and will continue to make reasonable efforts to correct inaccuracies, no warranty of accuracy is given. We invite all users to report any comments or issues via email to [email protected] (content) or [email protected] (privacy-related). The privacy policy governing this dataset is available at: https://www.fzi.de/en/privacy-policy/. For any public use or redistribution of the dataset, you must include proper attribution to the Licensor, provide a link to the Licensor's website, and license any adaptations under the same CC BY-NC-SA 4.0 terms. If this work proves helpful to your research, please consider citing the following BibTeX entry: @inproceedingspolley2025atlas, title=The ATLAS of Traffic Lights: A Reliable Perception Framework for Autonomous Driving, author=Polley, Rupert and Polley, Nikolai and Heid, Dominik and Heinrich, Marc and Ochs, Sven and Z"ollner, J. Marius, booktitle=Intelligent Vehicles Symposium (IV), organization=IEEE, year=2025

Cited by

• Fool the Stoplight: Realistic Adversarial Patch Attacks on Traffic Light Detectors

Related

• Traffic Light Recognition using Convolutional Neural Networks: A Survey
• TigAug: Data Augmentation for Testing Traffic Light Detection in Autonomous Driving Systems
• VATLD: A Visual Analytics System to Assess, Understand and Improve Traffic Light Detection
• Robust Traffic Light Detection Using Salience-Sensitive Loss: Computational Framework and Evaluations
• Patterns of Vehicle Lights: Addressing Complexities in Curation and Annotation of Camera-Based Vehicle Light Datasets and Metrics