Occ-MoE: Occlusion-Aware Mixture of Experts for Occluded Person Re-Identification

Abstract Person re-identification (ReID) suffers severely from occlusion, which corrupts pedestrian features and compromises the robustness of existing models. To tackle this, we propose a novel Occlusion-Aware Mixture-of-Experts (Occ-MoE) framework that introduces a "routing-and-restoration" paradigm for occluded ReID. Unlike conventional methods, our model features a Dynamic Occlusion-Aware Router (OAR) that adaptively guides information flow through experts based on occlusion patterns, effectively suppressing contaminated features. Critically, an Expert-based Restoration Module (ERM) is pioneered to recover identity-related cues from occluded regions by dynamically interacting router decisions with a dedicated occlusion stream. Extensive experiments on benchmark datasets demonstrate that our approach achieves new state-of-the-art performance, significantly improving both mAP and Rank-1 accuracy. This validates the framework's capability to mitigate feature corruption and actively restore discriminative information under severe occlusion.