TAU-106K: A New Dataset for Comprehensive Understanding of Traffic Accident

Authors: Yixuan Zhou, Long Bai, Sijia Cai, Bing Deng, Xing Xu, Heng Tao Shen

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate TABot s superior performance in traffic accident understanding, highlighting not only its capabilities in high-level anomaly comprehension but also the robustness of the TAU-106K benchmark. Our code and data will be available at https://github.com/cool-xuan/TABot.
Researcher Affiliation Collaboration 1 Alibaba Cloud Computing 2 Tongji University (Tongji)
Pseudocode Yes question_templates = [ "What's happened during {} in the video?".format, "What's the incident in the period of {}?".format, "Maybe something wrong happened during {} in the provided video?".format, "What's the traffic situation in the period of {}?".format, "Is the traffic flow captured by the video normal during {}?".format, "Dose the accident happen during {} in the video?".format, "Does the video record any traffic disruptions or accidents around {}?".format, "Is there any indication of an abnormal traffic event during {}?".format, "Could you identify any mishaps in the time frame of {}?".format, "Are there signs of vehicular distress or accidents within {}?".format ] User: random.choice(question_templates)(annotation["accident_segments"]) <video> GPT: annotation['accident_description']
Open Source Code Yes Our code and data will be available at https://github.com/cool-xuan/TABot.
Open Datasets Yes Our code and data will be available at https://github.com/cool-xuan/TABot.
Dataset Splits Yes For evaluation purposes, the TAU-106K dataset was split into training and testing sets in a 9:1 ratio, ensuring the same distribution of normal/accident instances and scene continuity across both.
Hardware Specification Yes We train LLM and both visual adapters of the Grounding GPT model through our TAU-106K dataset for 3 epochs using 8 H800 GPUs.
Software Dependencies No The paper mentions specific MLLMs used (Grounding GPT-7B, LLaMA70B) but does not provide specific version numbers for underlying software dependencies like programming languages or libraries (e.g., Python, PyTorch, TensorFlow, CUDA).
Experiment Setup Yes We train LLM and both visual adapters of the Grounding GPT model through our TAU-106K dataset for 3 epochs using 8 H800 GPUs. The initial learning rate is set to 2e-5 with a batch size of 32, requiring about 20 hours to complete.