Weakly supervised video anomaly detection with temporal attention module

Weakly supervised anomaly detection has become more attractive in terms of various application, such as surveillance video, and monitoring unacceptable video contents. A lot of existing weakly supervised learning methods depend on a Multiple Instance Learning (MIL) framework since MIL allows that a neural network yields temporal predictions when video level annotations are only given. Although those methods have shown outstanding performance, it has still problems in terms of the fact that they do not consider temporal dependencies among instances. To tackle this issue, we propose Weakly supervised video anomaly detection with temporal attention module, which facilitate the model to learn the relations of consecutive snippets of videos. In addition, we select top-k features in both abnormal segments and normal segments to maximize the separability of abnormal and normal instances. With the help of the proposed method, we improve performance on anomaly detection in UCF Crime dataset.