Domain-adaptive person re-identification via domain alignment and mutual pseudo-label refinement

Unsupervised domain-adaptive person re-identification refers to transferring knowledge from labeled to unlabeled datasets, thus alleviating the need for large amounts of labeled data. Existing methods address this problem using clustering methods to generate pseudo-labels. However, the pseudo-labels generated by current existing methods may be unstable and noisy, which will significantly degrade the performance of the method. In this paper, we propose a novel domain-adaptive person re-identification method via domain alignment and mutual pseudo-label refinement. First, we extract discriminative feature from the augmented data using a two-branch structure to enrich the feature diversity; second, we design a distributed adversarial domain alignment module to minimize domain differences; finally, we propose a consistency between local features and global features to refine pseudo-labels predicted by global features to exploit the complementary relationship between local and global features, and thus the noise generated by pseudo-label clustering is effectively reduced. Extensive experiments demonstrate that the proposed method can achieve remarkable results on popular benchmark datasets for domain-adaptive person re-identification.