Discriminative and Selective Pseudo-Labeling for Domain Adaptation

Unsupervised domain adaptation aims to transfer the knowledge of source domain to a related but not labeled target domain. Due to the lack of label information of target domain, most existing methods train a weak classifier and directly apply to pseudo-labeling which may downgrade adaptation performance. To address this problem, in this paper, we propose a novel discriminative and selective pseudo-labeling (DSPL) method for domain adaptation. Specifically, we first match the marginal distributions of two domains and increase inter-class distance simultaneously. Then a feature transformation method is proposed to learn a low-dimensional transfer subspace which is discriminative enough. Finally, after data has formed good clusters, we introduce a structured prediction based selective pseudo-labeling strategy which is able to sufficiently exploit target data structure. We conduct extensive experiments on three popular visual datasets, demonstrating the good domian adaptation performance of our method.