Semantic Mining and Cross-Center Synergy for Cross-Modal Person Re-Identification

Visible-Infrared Person Re-Identification (VI-ReID) involves matching pedestrian images across visible and infrared modalities, which faces a substantial modality discrepancy due to their distinct imaging principles. While existing methods attempt to reduce this discrepancy through image translation or feature alignment, they often overlook high-level semantic consistency and insufficiently explore intra-class and inter-class relationships for cross-modal metric learning. To address these issues, we propose a novel Semantic Mining and Cross-Center Synergy (SMCCS) network. Our framework introduces three key components: a Semantic-Aware Interactive (SAI) module that adaptively focuses on identity-relevant high-level semantic regions across modalities; a Fine-Grained Cross-Alignment (FGCA) loss that encourages a compact intra-class distribution in the shared feature space; and an Enhanced Cross-Triplet (ECT) loss that improves inter-class discriminability of cross-modal representations. Extensive experiments on two standard datasets, SYSU-MM01 and RegDB, demonstrate that our method achieves the state-of-the-art performance. Our code is available at https://github.com/GuYan66/ICASSP2026-SMCCS.