Unsupervised Person Re-identification Based on Camera-aware Guided Hybrid Contrastive Learning

Abstract

Unsupervised person re-identification has attracted more and more attention because of its extensive practical application prospects. Most of the existing methods are based on clustering technology to generate pseudo-labels to train models, ignoring intra-class differences caused by camera style gaps. In order to solve this problem, we propose a camera-aware contrastive learning method based on a camera-aware memory bank. The camera-aware memory bank stores the unique feature representation in each cluster with the same camera labels. Then intra-camera and inter-camera contrastive losses are designed to improve the model's robustness to camera styles. Experimental results show the effectiveness of our method. For two memtrics (mAP, Rank-1), our method achieves performance of 84.1%, 93.6% on Market1501, 40.7%, 71.2% on MSMT17 and 88.5%, 95.8% on PersonX, which outperforms other state-of-the-art methods.