O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

Researchers at Carnegie Mellon University introduce O-VAD, a training-free industrial video anomaly detector that replaces whole-frame classification with object-centric state evolution. Its three-stage pipeline uses GPT-5 from OpenAI to discover objects, SAM3 for concept-guided masks, CropFormer and SAM2 to build and recover spatiotemporal tubelets, and VLM queries to describe open-ended changes such as deformation, leakage, surface damage, and material release. A six-step reasoning chain then compares observed trajectories with expected process physics, assigns anomaly types and severity, localizes affected objects and frames, and applies visual verification to reduce false positives. Across Phys-AD, LiquidAD, and IPAD, O-VAD reaches a video-level AUROC of 0.584 on Phys-AD, exceeding GPT-5 at 0.503 and Qwen3-VL-32B at 0.513 without fine-tuning, domain knowledge, or predefined taxonomies. It also achieves 0.692 video-level AUROC on LiquidAD and 0.565 on IPAD, where anomalies involve multiple pipettes, periodic factory operations, and reference-based deviations. The ablation results identify state tracking as essential: removing it drives precision, recall, and F1 to 0 on 3 of 4 Phys-AD subsets. O-VAD produces interpretable reports that connect object states to causal diagnoses, although its multi-stage design adds latency and it remains weak on defects defined by invisible specifications, such as magnetic strength or exact actuation limits.