Dual Distillation Fusion for Weakly Supervised Anomaly Detection in Surveillance Videos

Anomaly detection in surveillance videos aims to differentiate anomalies from regular events by discriminative representations, which has gathered considerable attention due to its significant effect to public security. However, most existing works are limited in the lack of annotated samples, and lots of approaches find it challenging to avoid the well-reconstruction of anomalous data. To alleviate these issues, we propose a dual distillation fusion framework for weakly supervised anomaly detection. We reformulate the anomaly detection problem into two steps, namely filtering anomalies and inpainting normal patterns. Each step corresponds to one branch of the dual distillation. Specifically, the dual distillation comprises the contrastive distillation module and the inpainting distillation module. The contrastive distillation optimizes the encoder to filter out abnormal features and capture key normal features, while the inpainting distillation refines the decoder to inpaint normal patterns on the encoded features. The contrastive distillation module and the inpainting distillation module are optimized iteratively in a self-training manner with video-level labeled data. Moreover, a joint optimization module is devised to effectively fuse the distilled encoder and decoder, thereby collectively improving the anomaly detection performance. During the training phase, we take into account the diversity of normal samples by selecting pseudo normal and abnormal samples with high confidence from abnormal videos. These selected samples, along with original normal frames, are then fed into the subsequent training iterations to enhance the distinguishing ability of the model. Experimental results show that our proposed method performs competitively on five benchmark datasets.