Transformer-based Decoupled Modality Feature Learning for Visible-Infrared Person Re-Identification

Visible-Infrared person re-identification (VI-ReID) aims to match pedestrian images across modalities, requiring the simultaneous handling of intra- and cross-modality discrepancies. Existing dual-stream networks extract modality-specific features but often suffer from over-coupling and insufficient shared identity modeling. Simple feature fusion strategies do not adequately address the modality gap. We propose a ViT-based deep learning framework, termed Transformer-based Decoupled Modality Feature Learning (TDMFL), which effectively learns both modality-specific and modality-shared features while leveraging modality-invariant identity information to decouple different modality representations. Specifically, we first introduce an identity-modality decoupling learning strategy (IMDL) to facilitate learning with reliable modality-shared features while preserving essential modality-specific information. Additionally, we design a novel Identity-Modality Aggregation (IMA) loss function that efficiently integrates modality-specific and modality-shared features, assisting the model in learning more modality-invariant representations from both identity consistency and modality adaptation perspectives. Extensive experiments on the SYSU-MM01, RegDB, and LLCM datasets demonstrate that our method significantly outperforms existing state-of-the-art approaches. Code: https://github.com/hulu88/TDMFL.