WSAD-Net: Weakly Supervised Anomaly Detection in Untrimmed Surveillance Videos

Weakly supervised anomaly detection (WSAD) is a newfangled and challenging task, the goal of which is to detect anomalous activities in untrimmed surveillance videos with no requirement of temporal localization annotations. A few methods have been proposed to detect anomalies under the weakly supervised setting. In order to combat the issue even further, we present a weakly supervised anomaly detector network (WSAD-Net), which is composed of a pre-trained feature extractor and an anomaly-specific subnetwork. To learn the anomaly-specific parameters of WSAD-Net, we design a Classification Loss based on the multiple instance learning (MIL) and two novel losses, namely, Compactness Loss and Magnetism Loss, which play an important role in evaluating the correlation of features. During the test phase, we introduce the Anomaly-specific Temporal Class Activation Sequence (Ano-TCAS) to generate the anomaly score. We evaluate WSAD-Net on two benchmarks, i.e., the UCF-Crime and Live-Videos datasets, and experiments on these two benchmarks show that WSAD-Net outperforms or competes with current state-of-the-art methods.