Unsupervised person re-identification based on high-quality pseudo labels

The unsupervised domain adaptive (UDA) person re-identification (re-ID) method is of great significance to promote the practical application of person re-ID. However, the noisy pseudo labels in the target domain hinder its performance. In this paper, a novel high-quality pseudo labels (HQP) method for UDA person re-ID is proposed, which improves the performance from the perspectives of sample feature expression and similarity measurement in the clustering. In order to obtain better feature representation for target domain samples, a source domain generalization method based on contrastive learning (SCL) is designed. SCL learns the inherently consistent information within a sample, thereby improving the expression ability of the source domain pre-trained model. In order to provide a more reasonable similarity measurement for the clustering method, a soft label similarity based on neighborhood information integration (NII) is designed, which aids the clustering method to generate reliable pseudo labels. Market-1501, DukeMTMC-ReID and MSMT17 datasets are employed to evaluate the performance of the proposed HQP method. It achieves the results of 80.3%/92.3%, 68.0%/82.6% and 25.4/53.3 mAP/Rank-1 on DukeMTMC-ReID-to-Market-1501, Market-1501-to-DukeMTMC-ReID and DukeMTMC-ReID-to-MSMT17 tasks. Experimental results demonstrate that our HQP method performs favorably against the state-of-the-art UDA person re-ID methods.