O-VAD brings object-centric tracking and multi-agent reasoning to industrial video anomaly detection

arXiv·medium signal

O-VAD (arXiv 2607.18142) tackles industrial video anomaly detection by decomposing scenes into tracked objects and reasoning over them, rather than treating frames as monolithic inputs — the paper is filed under cs.MA alongside CV and CL, reflecting a multi-agent decomposition of the detection task. The approach targets identifying anomalous objects and events in manufacturing processes where a single global classifier loses the object-level provenance operators need. It is a concrete instance of agent decomposition improving an established vision task rather than replacing it.