Occluded Person Re-Identification With Pose Estimation Correction and Feature Reconstruction

In the real-world surveillance system application, accuracy of person re-identification (Re-ID) still suffers from occlusion. Occluded Re-ID aims to retrieve occluded images multiple cameras that do not overlap. To address this issue, we present a multi-branch feature enhancement network in paper. It has pose correction module (PCM), reconstruction (FRM), and part align (PAM). PCM builds key-points confidence corrected mechanism self-adaptive weights learning before matching stage strengthen part-level features non-occluded region weaken ones region. FRM reconstructs distribution inter-domain class-domain using maximum mean discrepancy generalize global features. The proposed separation enables two branches attend distinct pedestrian feature. increase power representation, PAM combines an outer product. experimental results on three holistic datasets, partial datasets demonstrate our method is superior. significantly outperforms state-of-the-art by 5.7% 4.4% rank-1 scores Occluded-Duke Occluded-REID, respectively.