Shao-Yuan Lo

Introduces a specialist visual assistant, Anomaly-OV, leveraging an anomaly expert and visual token selection mechanism to improve zero-shot anomaly detection and reasoning, establishing new datasets and benchmarks in the domain.

Proposes a rule-based reasoning framework, AnomalyRuler, for video anomaly detection using large language models, enabling fast scenario adaptation with few-normal-shot prompting and enhanced robustness through strategic modules.