Weakly-supervised Video Anomaly Detection

Weakly-supervised Video Anomaly Detection (WS-VAD) refers to identifying unusual or anomalous behaviors within video sequences using models trained primarily on video-level labels, without explicit frame-level annotations. Unlike fully-supervised methods, weakly-supervised approaches significantly reduce annotation costs by leveraging coarse labels (e.g., videos labeled as normal or anomalous). The primary challenge of this task is accurately localizing temporal anomalies and effectively distinguishing subtle anomalous activities from normal background events, relying only on limited supervision signals.