O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning
O-VAD claims a training-free way to spot industrial video anomalies by tracking how each object’s state changes over time.
The paper targets factory settings where object transformations, physics, and procedures make general VLM anomaly reasoning weaker. Its framework follows detected objects through spatial and temporal changes, then reasons over those object-level trajectories to identify abnormal objects in specific frames. The authors say it avoids retraining on normal clips and does not require domain-specific context at inference time. They report stronger results than frontier VLMs, agentic frameworks, and fine-tuned traditional VAD methods across three IVAD datasets, with interpretable anomaly reports. HF Daily Papers' note
The paper targets factory settings where object transformations, physics, and procedures make general VLM anomaly reasoning weaker. Its framework follows detected objects through spatial and temporal changes, then reasons over those object-level trajectories to identify abnormal objects in specific frames. The authors say it avoids retraining on normal clips and does not require domain-specific context at inference time. They report stronger results than frontier VLMs, agentic frameworks, and fine-tuned traditional VAD methods across three IVAD datasets, with interpretable anomaly reports. HF Daily Papers' note
score 4