wsjw233/CATSANet: v1.0.0CATSANet

We propose a Cross-Modal Semantic Token Selection and Alignment Network (CATSANet) that improves text-to-image person ReID by filtering identity-irrelevant visual tokens and enforcing fine-grained part alignment. Specifically, CATSANet contains three key components: (1) a cross-image semantic consistency-driven token selection module that suppresses noisy patches, (2) an Optimal Transport-based Part-Alignment Contrastive Loss (PACL) that performs soft, globally consistent alignment of modality-specific prototypes, and (3) a post-alignment residual interaction fusion module that adaptively integrates aligned local cues with global features to reduce representation gaps. Extensive experiments on CUHK-PEDES and RSTPReid show that CATSANet achieves competitive Rank-k and mAP performance while delivering more reliable fine-grained alignment and ranking quality across datasets.