Context-Aware Pseudo-Label Generation for Weakly Supervised Video Anomaly Detection

Weakly supervised detection of anomalies in surveillance videos seeks to identify abnormal events with only video annotations. Current approaches mainly focus on improving model performance by producing high-quality pseudo-labels based on snippet-level. However, the coarse-grained separability ensured by relying solely on video-level labels in pseudo-label-based methods neglects the fine-grained discriminability between different types of anomalies. In this paper, we propose a knowledge propagation framework based on text matching for video anomaly detection. Specifically, we first introduce a text matching mechanism that can identify video snippets associated with anomalous categories by utilizing anomaly textual prompts describing video anomaly types, thereby strengthening the model's ability to distinguish among different kinds of abnormalities. Furthermore, we devise a pseudo-label generation module leveraging knowledge propagation, which propagates representative knowledge to enhance the accuracy of generated pseudo-labels. Comprehensive experimental results demonstrate that our method surpasses current state-of-the-art techniques on the UCF-Crime and XD-Violence benchmark datasets.