MGFN++: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection

Weakly supervised anomalies detection in surveillance videos is a challenging task. The limitations of existing methods in localizing anomalies within lengthy videos include underestimating the high diversity of anomaly behaviors, misinterpreting high-speed motion with an anomaly, and overlooking the impact of scene variations. To overcome these challenges, we introduce a novel Glance and Focus Network that adeptly combines spatial–temporal information for precise anomaly detection. In the proposed framework, we incorporate a novel Dual−Behavior Prototype Memory Bank to learn the prototypes of anomaly behaviors and a novel Motion Decorrelation Mechanism to downplay features due to high-speed motion. In addition, we empirically found that current methods heavily rely on feature magnitudes to represent anomaly degrees, and inconsistent feature magnitudes across different scenes would easily confuse the neural network. To tackle this issue, we introduce the Feature Amplification Mechanism and a Magnitude Contrastive Loss, aimed at improving the discriminativeness of feature magnitudes for anomaly detection. Experimental results on two large-scale benchmarks, UCF-Crime and XD-Violence, demonstrate that the proposed MGFN++ surpasses state-of-the-art approaches.