A deep clustering-guide learning for unsupervised person re-identification

Abstract

Unsupervised person re-identification (RE-ID) has attracted increasing attentions due to its ability to overcome the scalability problem of supervised RE-ID methods. However, it is hard to learn discriminative features without pairwise labels and identity information in unlabeled target domains. To address this problem, we propose a deep clustering-guided model for unsupervised RE-ID that focuses on full mining of supervisions and a complete usage of the mined information. Specifically, we cluster person images from unlabeled target and labeled auxiliary datasets together. On the one hand, although the clustering IDs of unlabeled person images could be directly used as pseudo-labels to supervise the whole model, we further develop a non-parametric softmax variant for cluster-level supervision. On the other hand, since clustering badly suffers from intra-person appearance variation and inter-person appearance similarity in the unlabeled domain, we propose a reliable and hard mining in both intra-cluster and inter-cluster. Concretely, labeled persons (auxiliary domain) in each cluster are used as comparators to learn comparing vectors for each unlabeled persons. Following the consistency of the visual feature similarity and the corresponding comparing vector similarity, we mine reliable positive and hard negative pairs in the intra-cluster, and reliable negative and hard positive pairs in the inter-cluster for unlabeled persons. Moreover, a weighted point-to-set triplet loss is employed to adaptively assign higher (lower) weights to reliable (hard) pairs, which is more robust and effective compared with the conventional triplet loss in unsupervised RE-ID. We train our model with these two losses jointly to learn discriminative features for unlabeled persons. Extensive experiments validate the superiority of the proposed method for unsupervised RE-ID.