Relabeling Abnormal Videos via Intra-Video Label Propagation for Weakly Supervised Video Anomaly Detection

Weakly supervised video anomaly detection is chal-lenging due to the lack of frame-level labels in abnormal videos. Previous work generally formulates it as a multiple instance learning (MIL) problem, in which abundant snippets especially including hard snippets in abnormal videos are underutilized, causing the training process to be dominated by redundant easy snippets. In this work, we introduce a semi-supervised learning perspective and aim to provide reliable snippet-level labels for abnormal videos to fully leverage them. Specifically, we propose a relabeling strategy by utilizing the Label Propagation Network (LPN) to propagate labels from confident snippets to hard snippets. Considering label propagation tends to be fragile and inaccurate in this case due to incorrect pseudo labels, we develop a feature relation module to enhance the intra-class compactness and inter-class dispersion, as well as a label separation module to enlarge the discrepancy of propagated labels. The overall model is designed to be concise and universal, which can be applied to existing weakly supervised methods to improve performance. Extensive experiments demonstrate that our method achieves state-of-the-art results on three benchmark datasets.