This directory groups the baseline implementations used in the ACCIDENT dataset paper.
- heuristic/ - classical baselines with runnable Python scripts and matching notebooks
- llm/ - VLM/LLM baselines for temporal, spatial, and classification experiments
- Start with heuristic/README.md for the quickest smoke test.
- Move to llm/README.md once dataset paths and local environment are confirmed.
Both baseline families accept the normalized dataset layout under dataset/real_videos/. The runnable scripts also accept dataset/ as long as it contains real_videos/.