Industrial Video Anomaly Detection (IVAD) is critical for modern manufacturing and quality control, but existing vision-language models (VLMs) struggle with the intricate object transformations and strict physics of industrial settings. A new paper introduces O-VAD, a training-free agentic framework that mimics human inspectors by tracking object state evolution over time.
O-VAD operates in three stages: it grounds objects using SAM3, tracks them through physical transformations with tubelets, and reasons over per-object state trajectories to produce open-ended anomaly reports. This approach overcomes the limitations of prior methods that require retraining on normal clips or injecting domain knowledge.
Extensive experiments on three IVAD datasets—Phys-AD, LiquidAD, and IPAD—show that O-VAD outperforms frontier VLMs like GPT-5 and Qwen3-VL-32B, as well as traditional VAD methods fine-tuned on the respective datasets. The framework also provides interpretable reports detailing anomaly processes and types.
The authors argue that frontier VLMs fail at IVAD not because they cannot reason, but because they lack object-level evidence. O-VAD bridges this gap by grounding and tracking objects, enabling robust anomaly detection without any training or domain-specific knowledge.