Generation and Alignment: Bridging Modality Gap for Visible-Infrared Person Re-Identification

Most existing approaches for visible-infrared person re-identification (VI-ReID) attempt to learn shared features from visible and infrared images by either using a unified deep learning model or projecting the two modalities into a common intermediate space before feature extraction. However, due to the substantial modality gap, these methods often struggle to extract effective and distinctive shared features directly or may lose valuable identity-related information during the projection process. To address these issues, we propose a novel Tri-modal Generation and Alignment Swin (TGAS) network based on the Swin Transformer V2 framework, aiming to achieve distinctive features learning and effective cross-modality alignment. Specifically, we design a Hybrid Modality Generation (HMG) module, which consists of two key components: 1) a lightweight generation network that creates stylistically balanced hybrid modality features from pairs of infrared and visible features for per identity, and 2) a modality aligner that contains an adversarial loss to align different modalities via modality compensation embeddings at the low-level feature extraction stage, as well as an identity alignment loss to align identity distributions at the high-level feature extraction stage. Additionally, we introduce a Component Token Detection (CTD) module that enhances feature discriminability while mitigating the noise introduced by the generation process. Extensive experiments demonstrate that the proposed method achieves superior or, at the very least, comparable performance to state-of-the-art methods on the SYSU-MM01, RegDB, and LLCM datasets.