C-CANet:Calibration-Based Cross-Modal Attention Network for Audio-Visual Event Localization

The task of Audio-Visual Event Localization (AVEL) aims to detect and locate the temporal segments of events through the interaction of audio and visual modalities. However, existing methods often suffer from redundant information interference during modality interactions, which limits the effectiveness of feature alignment and multimodal representation learning. To address this issue, we propose a novel Calibration-based Cross-modal Attention Network for Audio-Visual Event Localization. Specifically, we introduce a set of globally learnable tokens into the multihead attention mechanism of audio and visual modalities. These tokens act as intermediate bridges, actively aggregating and filtering critical information from multimodal features while effectively suppressing redundancy and irrelevant information. Through this approach, the model can achieve more efficient information exchange and alignment between audio and visual modalities, thus enhancing multimodal representation learning. Experimental results demonstrate that the proposed method significantly outperforms existing approaches on the challenge AVE dataset, especially in complex scenarios, exhibiting stronger robustness and generalization. Our research indicates that the introduction of learnable interaction tokens is an effective strategy for improving the efficiency and quality of information exchange in multimodal tasks, providing a new solution for audio-visual event localization.