Boosting Weakly Supervised Video Anomaly Detection with Generative Description

With the extensive deployment of surveillance cameras, Weakly Supervised Video Anomaly Detection (WSVAD) has attracted increasing attention in many fields. It significantly reduces the labeling cost by relying only on video-level labels for training, and shows important significance in practical applications. However, existing methods often depend on unimodal visual information, neglecting the rich semantic information embedded in video description text. To address this limitation, this paper proposes a novel framework: Generative Description Boosted Weakly Supervised Video Anomaly Detection (DBVAD). DBVAD leverages large vision language models as the knowledge engine to generate video descriptions, which are then utilized as semantic supervision signals to optimize visual features. The proposed DBVAD comprises several key components. First, the key event selection strategy is used to accurately select key frames from videos for subsequent description generation. Second, the temporal modeling module captures the multi-scale temporal dependencies within videos. Lastly, the semantic focus prompt calibrates visual representations using label texts, while the description boosted module achieves fine alignment between visual features and generated description text through contrastive learning, thereby enhancing the model’s semantic understanding of abnormal events. Experimental results indicate that DBVAD achieves superior performance on the large-scale UCF-Crime and XD-Violence datasets, thereby validating its effectiveness.