Learning with noisy labels method for unsupervised domain adaptive person re-identification

Unsupervised domain adaptive (UDA) person re-identification (re-ID) aims to adapt the model trained on a labeled source domain to an unlabeled target domain. For pseudo-label-based UDA methods, pseudo labels noise is the main problem for model degradation and the factors that cause noise are complex. In this paper, a novel learning with noisy labels (LNL) method for UDA person re-ID is proposed to address this problem by analyzing the noise data itself. LNL learns with noise data from two aspects, including noise correction and noise resistance. According to the idea of neighbor consistency, pseudo labels correction (PLC) based on sample similarity is designed to correct the noisy pseudo labels before training. In order to solve the problem of noise labels in deep learning, noise recognition based on similarity and confidence relationship (SACR) is designed. Then, an easy-to-hard model collaborative training (MCT) strategy is developed, which can resist noise during the training process and obtain a more robust training model. To further avoid overfitting of noisy samples, the re-weighting (RW) method is employed in MCT. The proposed LNL model achieves considerable results of 75.2%/88.9% and 62.5%/77.4% mAP/ Rank-1 on DukeMTMC-reID-to-Market-1501 and Market-1501-to-DukeMTMC-reID UDA tasks. (c) 2021 Elsevier B.V. All rights reserved.