Refining Pseudo Labels for Unsupervised Domain Adaptive Person Re-Identification

Abstract

Unsupervised domain adaptive (UDA) person re-identification (re-ID) aims to generalize the model trained on a labeled source domain to an unlabeled target domain. Recently, the methods based on pseudo labels have achieved great success in the field of UDA person re-ID. However, pseudo label noise is still a problem to be solved. In this paper, we propose a novel framework to refine pseudo labels for UDA person re-ID. Firstly, in order to learn more discriminative features for better clustering, we propose a No-Reference Light Enhancement network (NRLE-Net) and a novel feature extraction model called Trans-Encoder. The NRLE-Net can reduce the inter-domain and intra-domain person image style differences caused by light condition changes simultaneously, which is helpful to improve the performance of the pre-trained model in target domain and reduce the noise of initial pseudo labels. The feature extraction model Trans-Encoder is designed based on Transformer. Compared with traditional Convolutional Neural Networks (CNN), the features extracted by Trans-Encoder are more robust to domain shift. On this base, the confidence of feature clustering can be improved. Secondly, to further reduce the noise of clustering pseudo labels, we design two noise refinement strategies. One is noise correction based on neighborhood consistency and the other is noise restraint based on label confidence. Through these two strategies, our model can be effectively optimized with the noise pseudo labels. Experimental results on Market-1501, DukeMTMC-reID and MSMT17 datasets demonstrate the effectiveness of our proposed UDA person re-ID framework.