Unsupervised domain specificity for knowledge transfer

Domain adaptation (DA) aligns domain invariant features between different domains as much as possible, but domain-specific features are largely ignored. In this paper, our aim is to utilize specific information on the target domain to improve domain adaptation. To address this issue, a natural idea is to use another classifier to model-specific information independently, because it is difficult to combine the different features of the two domains in a single model. Specifically, we propose a two-branch network (TBN), which utilizes ground truth labels to train the source domain classifier and utilizes pseudo-labels to train the target domain classifier. Different classifiers were able to learn different domain-specific features. Considering the noise impact of pseudo-labels, we propose one-time clustering module to further boost the accuracy of pseudo-labels. In particular, TBN can be easily integrated into various DA methods to further improve their performance. The superiority of the proposed method is validated on several standard datasets.