Mutual Learning Inspired Prediction Network for Video Anomaly Detection

Video anomaly detection has made great achievements in security work. A basic assumption is that the abnormal is the outlier of the normal. However, most existing methods only focus on minimizing the reconstruction or prediction error of normal samples while ignoring to maximize that of abnormal samples. The completeness of the training data and the similarity between certain normal and abnormal samples can cause the network overfitting to normal samples and generalizing to abnormal samples. To address the two problems, we propose Mutual Learning Inspired Prediction Network. Specifically, it consists of two student generators and one discriminator to predict the future frame, together with our proposed Boundary Perception-Based Mimicry Loss and Self-Supervised Weighted Loss. The proposed Boundary PerceptionBased Mimicry Loss guides the generators to learn the predicted frame from each, which can help to increase the diversity of training data and prevent interference at the same time. The proposed Self-Supervised Weighted Loss constraints the confusion samples in training data with a small weight, which can clarify the modeling goal of the network and enlarge the distance between normal and abnormal samples. Experiments on four mainstream datasets demonstrate the effectiveness of our proposed method.