Cross-domain person re-identification via learning Heterogeneous Pseudo Labels

Abstract

Assigning pseudo labels is vital for cross-domain person re-identification (ReID), and most existing methods only assign one kind of pseudo labels to unlabeled target domain samples, which cannot describe these unlabeled samples accurately due to large intra-class and small inter-class variances caused by diverse environmental factors, such as occlusions, illuminations, viewpoints, and poses, etc. In this paper, we propose a novel label learning method named Heterogeneous Pseudo Labels (HPL) for cross-domain person ReID, which could overcome large intra-class and small inter-class variances between pedestrian images in the target domain. For each unlabeled target domain sample, HPL simultaneously learns three different kinds of pseudo labels, i.e., fine-grained labels, coarse-grained labels, and instance labels. With the three kinds of labels, we could make full use of their own advantages to describe target domain samples from different perspectives. Meanwhile, we propose the Pseudo Labels Constraint (PLC) to improve the quality of the heterogeneous labels by using their consistency. Furthermore, in order to relieve the influence of noisy labels from the aspect of contrastive learning, we propose the Confidence Contrastive Loss (CCL) to consider the sample confidence in the learning process. Extensive experiments on four cross-domain tasks demonstrate that the proposed method achieves a new state-of-the-art performance, for example, the proposed method achieves 87.2% mAP and 95.0% Rank-1 accuracy on MSMT17$\to$Market.