Modality-Parallel Disentanglement and Contrastive Optimization for Efficient Visible-Infrared Person Re-Identification

Visible-Infrared Person Re-Identification (VI-ReID) addresses the challenging task of matching individuals across visible and infrared camera modalities. However, existing classification-based methods often suffer from limited generalization capability due to residual modality-specific features. To address this challenge, we introduce a modality-parallel learning framework that simultaneously processes and disentangles cross-modal representations through two novel components: Modality-Parallel Representation Disentanglement (MPRD) and Dual Contrastive Separation (DCS). The MPRD module operates in a modality-parallel manner to suppress identity-irrelevant variations while enhancing discriminative features in a unified embedding space. Complementarily, the DCS strategy further reinforces the separation of identity-relevant information from modality-specific factors, promoting better alignment and disentanglement. Experiments on the SYSU-MM01 and RegDB datasets show that our method achieves leading performance, attaining a Rank-1 accuracy of 74.99% and an mAP of 68.84% under the all-search mode on SYSU-MM01, which significantly surpasses current state-of-the-art methods.