ABPAT : Auxiliary Branch and Patch Attention Transformer for Occluded Person Re‐Identification

The task of occluded person re‐identification aims to address the issue of matching severely occluded person images. Occlusion not only fragments and reduces visible identity cues, but also introduces occluder and background noise, resulting in feature contamination. Conventional person re‐identification methods often fail to achieve satisfactory performance in occluded person re‐identification tasks. To address these challenges, we propose a novel Auxiliary Branch and Patch Attention Transformer (ABPAT). Specifically, ABPAT is composed of four components: auxiliary branch, attention branch, Relevance Score‐based Patch Attention Module (RSPAM), and Two‐Step Feature Fusion Module (TSFFM). The ViT attention branch models global relationships, while the CNN auxiliary branch complements fine‐grained local information. RSPAM reweights the local features of the attention branch based on spatial dependencies among patches and cosine relevance between each local patch and the holistic feature, thereby reducing occlusion interference. Meanwhile, TSFFM integrates the features from both branches in two steps by first fusing their local features and then combining them with the holistic token. On Occluded‐DukeMTMC, ABPAT achieves 62.2% mAP and 72.9% Rank‐1, the best results among the compared methods, and it also obtains competitive performance on Occluded‐REID and the holistic Market‐1501 and DukeMTMC‐ReID datasets. Extensive experiments conducted on both occluded and holistic person re‐identification datasets and further analyses of the modules, occlusion robustness, feature distributions, and inference efficiency validate the effectiveness of the proposed method and its practicality.