Mixture-of-Experts Vision Transformer for Occluded Person Re-Identification

Abstract Person re-identification (ReID) is crucial for cross-camera target retrieval, yet it faces significant challenges due to occlusion. We introduce the Occlusion-Aware Mixture-of-Experts Vision Transformer (Occ-MoE), a dual-branch architecture that processes both original and occluded images. Occ-MoE incorporates an Occlusion-Aware Routing (OAR) mechanism and a Dual-Path Experts (DPE) module to emphasize robust identity-related features while suppressing occluded or non-pedestrian regions. Our Expert-based Restoration Module (ERM) facilitates dynamic interaction between occlusion stream features and predicted occlusion masks. Experiments on several benchmark datasets demonstrate superior ReID performance, improving both mean Average Precision (mAP) and Rank-1 accuracy, validating our design's effectiveness in mitigating feature corruption and performing occlusion feature recovery. Code is available at: https://github.com/BestTao/occ-moe.