Unsupervised person re-identification by dynamic hybrid contrastive learning

Abstract

Unsupervised person re-identification (re-ID) aims at cross-camera pedestrian retrieval without manual annotation. Recently, the contrastive learning has been introduced into the field of unsupervised person re-ID. However, existing methods usually focus only on mining the intra-category similarity and neglect the negative effects of clustering noise, which limits the performance in unsupervised person re-ID. In this paper, we propose a Dynamic Hybrid Contrastive Learning (DHCL) method for unsupervised person re-ID. Specifically, we perform the clustering algorithm and the dynamic refinement policy to divide the unlabeled training dataset into two subsets, i.e., clustered samples with pseudo labels and un-clustered independent instances, at each training epoch. Then, the proposed DHCL guides the feature extraction network to mine the intra-category similarity from the clustered samples by applying attraction within samples of the same cluster. Meanwhile, the inter-instance discrimination is also mined by pushing away different instances. Besides, we integrate the two levels of contrastive learning into an end-to-end framework and exploit the complementarity between them to improve the separability of the feature space. To reduce the negative effect of over-focusing on positive samples, a penalty item is added to the hybrid contrastive loss. Extensive experiments demonstrate the effectiveness of the proposed method in unsupervised person re-ID.