MTC-Net: Multi-granular Temporal Cascade Learning for Weakly Supervised Video Anomaly Detection

In this paper, we propose a novel Multi-granular Temporal Cascade Learning framework (MTC-Net), which effectively combines coarse-grained and fine-grained detection capabilities. By leveraging visual features extracted from CLIP and integrating visual-language associations, MTC-Net addresses the limitations of existing methods and achieves outstanding performance across multiple benchmark datasets. Furthermore, we incorporate multiscale temporal feature fusion and an improved attention mechanism into the framework, and introduce a feature consistency-based auxiliary loss to enhance the model's focus on anomalous patterns and generalization ability. Extensive experiments on the XD-Violence and UCF-Crime datasets demonstrate that MTC-Net outperforms both semi-supervised and classification-based weakly supervised methods across all evaluation metrics. For the coarse-grained Weakly Supervised Video Anomaly Detection (WSVAD) task, MTC-Net achieves an AP of 84.83% on XD-Violence and an AUC of 75.23% on UCF-Crime, significantly surpassing current state-of-the-art methods. In the more challenging fine-grained WSVAD task, MTC-Net achieves a 13.1% improvement in average precision over previous methods. These results validate the effectiveness of our approach, highlighting MTC-Net's potential to advance the field of video anomaly detection through the efficient use of weak supervision, multiscale temporal feature fusion, and multimodal feature integration.