Traffic Anomaly Reasoning (TAR)

Summary

The official dataset for the AI City Challenge 2026 Track 3 (Anomalous Events in Transportation), containing 44,040 pseudo-labeled multi-task training annotations across 3,670 CCTV videos (~26.1 hours) sourced from eight public datasets, plus 960 human-curated test annotations from 80 clips trimmed from 17 public YouTube videos. Each video is annotated across 10 task types spanning basic QA, scene/video understanding, and temporal reasoning, with explicit chain-of-thought reasoning traces generated by a hierarchical auto-labeling pipeline using Gemini 3.1 Pro and Gemma-4. Videos are not redistributed; a download script fetches them from original sources (~150 GB total).

Keywords

video video-understanding anomaly-detection traffic surveillance vlm reasoning chain-of-thought cctv ai-city-challenge

Topics

Vision / Video Understanding / Anomaly Detection

Research notes

• Annotations generated by hierarchical auto-labeling pipeline (Gemini 3.1 Pro for captioning, Gemma-4 for Q&A generation). 910 of 3,670 videos had supplementary NVIDIA human annotations. Test split answers are redacted; predictions must be submitted to an evaluation server. Dataset is ready for commercial use.