Region Erasure Fusion Network for Occluded Person Re-Identification

Person re-identification(ReID) is an important task in computer vision. Although some random erasure preprocessing methods have made significant progress in sample representation, they often fail to erase important regions, resulting in the problem of insufficient occlusion adaptation. To address this problem, we propose a region erasure fusion network(IRE-ReID). Firstly, an important feature region erasure model is designed, which identifies the important feature regions through region quality prediction, and the regions are partially erased to form pedestrian occlusion samples, thereby improving the diversity of the samples of persons. Finally, the person-erase feature weighted fusion module is presented to identify the regions of interest from different erasure features through the fusion of person foreground features and erasure features. This process enhances the concentration on the region of differentiating pedestrians. Extensive experimental results have demonstrated that the proposed IRE-ReID achieves superior identification performance compared to several state-of-theart methods.