基于局部特征匹配和伪标签细化的纯无监督行人重识别OA

In allusion to the problem of significant noise in pseudo-labels generated by clustering in unsupervised person re-identification(ReID),a purely unsupervised method based on local feature matching and pseudo-label refinement is proposed.This method does not rely on any source domain information,but only considers the correlation between samples at the image level and assigns robust pseudo-labels for training.A local feature matching module is designed to align and rank local features of samples,so as to represent the correlation between global features and local features of samples reasonably.Then,a correlation scoring module is used to score the rationality of the generated pseudo-labels by considering the correlation between global features and local features comprehensively.On this basis,a pseudo-label refinement module is introduced to refine the pseudo-labels of global features and local features based on the scores of samples.The refined pseudo-labels are used to train the net-work and continuously update the pseudo-labels.The experimental verification of the method is conducted on the public per-son ReID datasets Market-1501,DukeMTMC-ReID and MSMT17.The results show that the mAP of this method can reach 81.9%,71.1%and 31.6%on the Market-1501,DukeMTMC-ReID and MSMT17 datasets,respectively,demonstrating better per-formance.