From Detection to Understanding – A Multi-Task Dataset for Traffic Anomaly Reasoning
Published in 2026 Conference on Neural Information Processing Systems, 2026
Recommended citation: Han Zhang, Yilin Zhao, Zaid Pervaiz Bhat, Zheng Tang, Varun Praveen, Vidya Nariyambut Murali, David C. Anastasiu and Tomasz Kornuta. "From Detection to Understanding -- A Multi-Task Dataset for Traffic Anomaly Reasoning". Proceedings of Conference on Neural Information Processing Systems (NeurIPS 2026). 2026. https://arxiv.org/abs/2608.10317
Abstract
TAR (Traffic Anomaly Reasoning) is a large-scale multi-task dataset for training and evaluating vision-language models on traffic anomaly understanding. It contains 44,040 training annotations with explicit chain-of-thought reasoning traces across 10 task types and 3,670 CCTV transportation videos (approximately 26 hours) sourced from eight public datasets. Its tasks span Question Answering, Temporal Reasoning, and Scene Understanding. Experiments on a human-reviewed held-out set show that current models struggle with deeper temporal and scene reasoning even when they perform well on basic question answering, while multi-task fine-tuning yields substantial gains. TAR is released under CC BY 4.0 and serves as the official training data for AI City Challenge 2026 Track 3.
BibTeX
@InProceedings{Zhang26TAR,
author = {Han Zhang and Yilin Zhao and Zaid Pervaiz Bhat and Zheng Tang and Varun Praveen and Vidya Nariyambut Murali and David C. Anastasiu and Tomasz Kornuta},
title = {From detection to understanding – a multi-task dataset for traffic anomaly reasoning},
booktitle = {Proc. NeurIPS},
address = {Atlanta, GA, USA},
year = {2026}
}