Enhancing Cross-Modality Person Re-Identification through Attention-Guided Asymmetric Feature Learning

The main objective of the visible-infrared person re-identification (VI-ReID) task is to correlate images of pedestrians captured in daylight with their corresponding infrared (IR) representations. The main challenge lied in the cross-modal and intra-modal differences between VIS and IR images, which often result in reduced recognition accuracy. To address these changes, this paper uses a middle modality generator(MMG) that converts pedestrian images into middle-modality(M-modality) ones, transforming the dual-modality task into a tri-modality one. Commonly, VI-ReID utilizes a dual-branch network to extract features from two modalities. However, to enhance feature extracition across three modalities, we propose a four-branch parameter sharing network(FBPN) and explore its parameter sharing capabilities. Research conducted on datasets showed that the FBPN effectively minimizes modality disparities and mitigates background interference through a lightweight channel attention module. Furthermore, we introduce an Asymmetric Multi-Granularity Feature Learning (AMFL) module to further decrease modality diversity. The proposed method achieves a 70.60\% mean Average Precision in the All-Search mode on SYSU-MM01 and 86.7\% mean Average Precision in the VtI mode on RegDB, outperforming existing approaches.