license: mit
language:
- en
Roundabout-TAU: A Real-World Traffic Anomaly Understanding Dataset
Overview
Roundabout-TAU is a real-world traffic anomaly dataset that support multiple tasks and mainly support Multi-Modality Large Language Models(MLLMs) traffic anomaly understanding. We collect 342 video clips, includes 276 anomaly clips and 66 normal clips, from 28 fixed roadside cameras through the cooperation with the City of Carmel, Indiana. We further provide test-based QA annotations to support training for MLLMs.
Dataset Contents
/videos.zip: contains all 342 videos./annotations/roundabout_vau_summary_qa_pairs.json: Anomaly summarization task for all anomalous videos. This requires the model to summarize the anomaly event from the perspectives of environment, object grounding, description, and reasoning, forming a coherent paragraph./annotations/roundabout_vau_category_qa_pairs.json: Anomaly classification/detection task for all videos. The dataset provides 4 anomaly category labels: A: No anomaly, B: Direction or maneuver violation, C: Near-collision or collision, D: Abnormal road use. Further definitions of these categories are provided in the system prompt./annotations/roundabout_vau_object_grounding_qa_pairs.json: Anomaly object grounding task for anomalous videos. This annotation includes environmental and frame-level grounding./annotations/roundabout_vau_time_window_qa_pairs.json: Anomaly temporal grounding task for anomalous videos. The task requires the model to generate a temporal window [start_sec, end_sec] indicating the duration of the anomaly event./annotations/roundabout_vau_simple_description_qa_pairs.json: Anomaly description task for anomalous videos. This requires the model to summarize the anomaly event in one simple sentence./annotations/roundabout_vau_simple_reasoning_qa_pairs.json: Anomaly reasoning task for anomalous videos. This requires the model to provide a simple explanation of why the anomaly event occurred./annotations/roundabout_vau_environment_qa_pairs.json: Environmental perception task for all videos. This asks the model to identify road type, weather, road surface, and time of day./annotations/roundabout_vau_all_qa_pairs.json: A collection of all annotated QA pairs.
Dataset Split Information
We randomly split dataset into training (300 videos) and test sets (42 videos) to test our TAU-R1 framework. The splits are documented in /annotations/roundabout_tau_train_videos.txt and /annotations/roundabout_tau_test_videos.txt. The corresponding JSON annotations for the train/test splits are in /splited_annotations.
How to Use
- download/extract all videos sources in corresponding folder.
- download task-specific json and change video path to your corresponding video directory.
More Details(Paper)
More details of this dataset can be got from paper TAU-R1.
License
This dataset is released under the MIT License. You are free to use, modify, and distribute this dataset in accordance with the terms of the MIT License.
See the LICENSE file for more details.
Citation
If you use this dataset in your research, please cite the following paper:
@misc{lin2026taur1visuallanguagemodel,
title = {TAU-R1: Visual Language Model for Traffic Anomaly Understanding},
author = {Yuqiang Lin and Kehua Chen and Sam Lockyer and Arjun Yadav and Mingxuan Sui and Shucheng Zhang and Yan Shi and Bingzhang Wang and Yuang Zhang and Markus Zarbock and Florain Stanek and Adrian Evans and Wenbin Li and Yinhai Wang and Nic Zhang},
year = {2026},
eprint = {2603.19098},
archivePrefix= {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2603.19098}
}