Short-Long Temporal Normality Deviation for Weakly Supervised Video Anomaly Detection

Weakly supervised video anomaly detection (WSVAD) tries to find abnormal events in untrimmed videos, while only video-level labels are given. However, many existing methods still have two main problems. First, abnormal events may last for different lengths of time, and their temporal changes are also not always the same; therefore, it is hard for one single time scale to collect enough useful anomaly information. Second, most current methods select abnormal segments mainly according to high anomaly scores, but these scores may be influenced by obvious visual regions, background changes, or scene-related bias, so the final anomaly localization may become unstable. To deal with these problems, this paper proposes STND-VAD, a short-long temporal normality-deviation framework for WSVAD. In this framework, a multi-scale dilated temporal feature enhancement module is first used to obtain contextual information from different temporal ranges. Then, this paper builds short-term and long-term MLP discriminative branches, so that local instant changes and longer temporal changes can be modeled separately. Besides, a normality-deviation evidence calibration mechanism is designed, which learns normal prototypes and measures how much each segment deviates from these normal patterns; in this way, it provides extra constraints for anomaly responses and reduces the influence of unreliable high-response segments. Experiments on three public benchmark datasets show that STND-VAD is effective and can achieve competitive anomaly detection results.