Weakly Supervised Video Anomaly Detection

Weakly supervised video anomaly detection uses labels for whole videos or clips to learn which moments, and sometimes which regions, contain unusual behavior. Common approaches include multiple-instance learning, temporal modeling at different scales, spatial patch scoring, and vision-language models that contribute semantic cues or pseudo-labels. These methods are often evaluated on surveillance benchmarks such as UCF-Crime, XD-Violence, and ShanghaiTech, where reliable localization could help analysts find events in long recordings. A central tension is that high pooled frame-level scores may reflect scene or recording cues rather than true within-video localization, while open-world detection also struggles with missed anomalies and sensitivity to prompts or context.