Hard-sample guided cluster refinement for unsupervised person re-identification

Unsupervised person re-identification (Re-ID) aims to learn robust and discriminative features with unlabeled data. Recently, more attention of clustering-based methods focus on using cluster centroids and all instances for contrastive learning. However, the previous methods did not fully consider the information of hard samples in clustering process and cluster-level contrastive learning process. In this paper, we propose a novel hard-sample guided cluster refinement (HGCR) approach to learn information of hard samples in a simple but efficient way. Specifically, in HGCR we improve the reliability of clustering-based pseudo labels under the guidance of the historical cluster information. Meanwhile, we introduce class-level and instance-level contrastive learning with a sample filter scheme, which can fully exploit the information of hard positive samples. Extensive experiments on three large-scale person Re-ID benchmarks demonstrate the effectiveness of the proposed method, which outperforms state-of-the-art unsupervised person Re-ID methods by a considerable margin.