Federated weakly-supervised video anomaly detection with mixture of local-to-global experts

Yong Su, Jiahang Li, Simin An, Meng Xing, Zhiyong Feng

Abstract: Highlights•Design a federated weakly supervised framework for privacy in multi-scene surveillance.•Create a mixture of experts model integrating local and global representations to handle domain shift.•Localize anomalous events using gate scores of local tubes, not just anomaly scores.•Define an anomaly localization metric that does not depend on auxiliary tasks.