## Motivation

Traffic accidents are rare but high-impact events. Being able to automatically understand **what happened and where it happened** can support faster incident response, better safety analytics, and more reliable reporting—especially in places where fixed CCTV cameras are the main source of coverage. But CCTV video is challenging: crashes can be far from the camera, partially blocked, low-resolution, and visually cluttered, which makes manual review slow and expensive.

Most video benchmarks assume you can train on a large labeled dataset from the same source. Real deployments often don’t have that luxury. Labels are hard to get, rules vary by location, and camera setups change over time. That’s why this competition is set up differently: **we do not provide a labeled real training split.** The goal is to test methods that generalize well **without** tuning on an official real training set.

To make the task practical, we provide **synthetic data** that you can use to prototype your pipeline, pretrain models, or validate ideas before running on the real test clips.

## What will you learn?

You’ll learn how to build a pipeline for **CCTV accident understanding** that predicts three things from each clip:
- **when** the accident happens (time),
- **where** it happens in the frame (location),
- and **what type** of collision it is.

You’ll also learn how to design and evaluate models in a **training-free / zero-shot** setting, where success depends on robustness and generalization rather than dataset-specific fitting.