Reproduce “Zero-Shot Traffic Accident Detection via a Coarse-to-Fine VLM-Tracking Pipeline” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Dipit Saha, Shah Mohammad Abdul Mannan, Mohammad Raihan Rashid, Ruwad Naswan, Ahnaf Tahmid
A training-free pipeline combines a vision-language model with object detection and tracking to perform zero-shot traffic accident detection, achieving a 0.504 three-way harmonic mean score and a 22% relative improvement over prior baselines on real-world CCTV footage.
Key insights
- Augmenting VLM inputs with structured visual information (annotated bounding boxes, explicit coordinate text) and temporal context significantly enhances localization accuracy and reduces reliance on raw pixel interpretation.
- A coarse-to-fine temporal strategy effectively mitigates VLM biases (e.g., mid-video prediction) and ensures critical moments are captured, even with challenging video characteristics.
- Implementing inference-time safeguards, such as YOLO retry, raw-frame fallback, and JSON recovery, is crucial for system robustness and preventing catastrophic failures in real-world zero-shot scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.