O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning
O-VAD tracks how each object changes over time, then uses those trajectories to flag abnormal industrial events.
The paper says industrial anomaly detection breaks down for general VLM methods when processes involve object transformations, physics rules, and procedural constraints. Its proposed framework is training-free and avoids domain-specific knowledge at test time. The authors report stronger results than frontier VLMs, agentic frameworks, and fine-tuned traditional video anomaly methods across three IVAD datasets. It was accepted to ECCV 2026. ArXiv · AI/CL/LG's note
The paper says industrial anomaly detection breaks down for general VLM methods when processes involve object transformations, physics rules, and procedural constraints. Its proposed framework is training-free and avoids domain-specific knowledge at test time. The authors report stronger results than frontier VLMs, agentic frameworks, and fine-tuned traditional video anomaly methods across three IVAD datasets. It was accepted to ECCV 2026. ArXiv · AI/CL/LG's note
score 4