Local-Aware Residual Attention Vision Transformer for Visible-Infrared Person Re-Identification

Visible-infrared person re-identification (VI-ReID) task is to retrieve the same pedestrian across the visible and infrared modalities. The existing transformer-based works are constrained by the inherent structure of the ViT that feature collapse in deeper layers and the over-globalization of extracted features, resulting in incomplete learning of local and low-level features. However, these features are instrumental in representing and identifying elements within visible-infrared images more comprehensively, which increases the accuracy and robustness of cross-modal pedestrian matching. To solve the above problem, we propose the Local-Aware Residual Attention Vision Transformer (LAReViT) to enhance the learning of fine-grained local and shallow-level information to reinforce the feature discrimination and comprehensiveness in ViT. Specifically, the Local-Aware Residual (LAR) Module, which uses a novel Local Residual Attention (LRA) mechanism, is proposed to increase the fine-grained local information contained in feature extraction. In order to exploit fine-grained local information lost in lower-level visual features, the LRA in the LAR module adopts novel attention residual connections. Additionally, we propose a Positional Channel Reconstruction (PCR) Module that takes advantage of the local receptive field benefits of convolution. PCR reweights features within patches at the channel level, further facilitating the network emphasis on effective fine-grained local information. Finally, the novel Center Aggregation Loss (CAL) is designed to reduce modality discrepancies moderately and promote comprehensive feature extraction. Extensive experiments conducted on the SYSU-MM01, RegDB, and LLCM datasets demonstrate the state-of-the-art performance achieved by our proposed method. The code is available at https://github.com/Hua-XC/LAReViT .