Answers are grounded in this paper and cite their sources.
Existing weakly supervised anomaly detection methods suffer from insufficient semantic alignment of features and a lack of fine-grained localisation capabilities. Furthermore, purely visual models have limitations in understanding semantic information, making it difficult to effectively integrate visual features with semantic information, which results in poor discrimination of multi-class anomalies. To address these issues, this paper proposes a vision-language aligned video anomaly detection model. By introducing a vision-language alignment branch and utilising a pre-trained CLIP model to achieve fine-grained semantic alignment between video features and category text prompts, this model overcomes the limitations of purely visual models in understanding semantic information.It designs an adaptive gated fusion mechanism to dynamically fuse the global anomaly scores from the original visual branch with the semantically guided scores from the alignment branch, combining the complementary strengths of visual pattern recognition and semantic understanding to enhance the model’s ability to distinguish between multiple anomaly categories.And it constructs a multi-task loss function to jointly optimise temporal localisation and fine-grained classification tasks, making full use of video-level weak supervision signals and cross-modal alignment information. Experimental results on the UCF-Crime and XD-Violence datasets demonstrate that this method effectively improves fine-grained anomaly localisation and classification performance, exhibiting significant advantages.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: