Occluded Person Re-identification Based on Pose-guided Feature Enhancement
To address the problems of feature loss during feature extraction and noise interference during feature matching in occluded person re-identification,a pose-guided feature enhancement model is proposed.Firstly,with the assistance of key point information,a symmetrical region feature repair module is proposed to replace the locally lost features in the occluded region with the features in the non-occluded region.Secondly,to explore the semantic relationship between local features,an adjacent region feature compensation module is proposed to update the feature representation of each local feature by combining the features of adjacent regions.Finally,by leveraging the generalized mean pooling to extract features from the central region of the feature map,the expression ability of person feature vectors is improved to obtain more accurate global features.Simulation experiments show that the proposed model outperforms most algorithms in common holistic datasets,partial datasets,and occluded datasets in terms of Rank-1 and mAP,achieving 56.7％and 72.4％Rank-1 on Occluded-Duke and Occluded-REID datasets,respectively.
person re-identificationocclusionfeature repairfeature compensationgeneralized mean pooling
万方数据
维普
