Visible-Infrared Person Re-Identification Network Based on Intermediate Feature Learning

Visible-infrared person re-identification (VI-ReID) aims to retrieve images of the same individual captured by visible and infrared cameras. With the advent of Vision Transformers (ViTs), existing methods have benefited from a strong global receptive field during the feature extraction stage.However, their channel attention mechanisms often lack flexibility and operate at a coarse granularity. Additionally, due to the significant modality gap between visible and infrared images, learning intermediate modality features remains a challenging task.To address these issues, we propose DFHNet, a novel network based on intermediate feature learning to facilitate effective VI-ReID. Specifically, we first employ a multi-modal feature module to extract both local and global features from visible and infrared images. Then, we introduce a cluster-guided contrastive fusion module, which leverages contrastive learning to integrate intermediate features generated from different modalities and scale sizes, effectively capturing view-common features that encapsulate both shared and complementary information.Extensive experiments on the SYSU-MM01 and RegDB datasets demonstrate that our method achieves state-of-the-art performance on the VI-ReID task.