Bidirectional modality information interaction for Visible-Infrared Person Re-identification

Due to the distinct imaging principles of visible and infrared images, significant challenges arise for visible- infrared person re-identification (VI-ReID). Current methods either establish intermediate modalities to reduce image differences or use various techniques to extract better shared features. However, despite the modality differences in VI-ReID, significant similarities in contours and postures exist, so focusing on learning feature representation can fundamentally solve the problem. To this end, this paper proposes a bidirectional modality information interaction network to reduce the difficulty of inter-modality recognition and enhance intramodality feature extraction. To fully exploit cross-modality information, the Dynamic Aggregation (DA) module iteratively projects features of both modalities onto each other, embedding image attributes from both, enabling mutual learning, improving feature representation, and reducing modality differences. To enhance intramodality feature representation, the Feature Intersection (FI) module combines two pooling styles in a novel way, ensuring global connection of local features and regional highlights. Lastly, we propose a Multi-Weight (MW) loss to increase the center distance of same identity features within the same modality, helping the model better learn identity attributes. Extensive experiments on SYSU-MM01 and RegDB confirm the superiority of our approach.