PhysicalAI-Traffic-Anomaly-Reasoning

Traffic Anomaly Reasoning (TAR) This dataset is the official release for **AI City Challenge 2026 Track 3 — Anomalous Events in Transportation**. It contains: **Training** (`train/`): 44,040 pseudo-labeled multi-task annotations covering 3,670 CCTV videos (≈26.1 hours: 9.2 hrs anomalous + 16.9 hrs normal) sourced from **eight public datasets**. **Test** (`test/`): 960 human-curated annotations covering **80 short clips trimmed from 17 public YouTube videos**. Answers are redacted in this release; submit predictions to the evaluation server to receive leaderboard scores (see Test Split). Each video is annotated across **10 task types** spanning basic question answering, scene/video understanding, and temporal reasoning, with explicit chain-of-thought reasoning traces accompanying every answer. **Videos are not redistributed in this release.** A download script is provided to fetch the source videos from their original public repositories. See Downloading the Source Videos below and the companion guide `DOWNLOADING.md` for full details. Every source — including SO-TAD — is fetched automatically. Quick Start Dataset Description TAR is built for training and evaluating video vision-language models (VLMs) and AI systems on tasks including anomaly detection, video event reasoning, temporal localization, causal understanding, and scene description in transportation monitoring contexts. Relevant applications include smart city traffic monitoring, road safety AI, and general video event understanding. Annotations were generated by a hierarchical auto-labeling pipeline: 1. **Three-level video captioning (Gemini 3.1 Pro)** — global scene caption, temporally grounded dense caption, and fine-grained chunk captions over short video segments. 2. **Multi-Scale Spatio-Temporal Event Description (MSTED) synthesis (Gemini 3.1 Pro)** — structured consolidation of caption levels into a holistic event representation, incorporating human annotations where available. 3. **Multi-task Q&A an