Pseudo-labeling Integrating Centers and Samples with Consistent Selection Mechanism for Unsupervised Domain Adaptation

Abstract

Pseudo-labeling is widely applied to generate pseudo labels of target samples in most Unsupervised Domain Adaptation (UDA) methods. Existing UDA methods designed the pseudo-labeling strategy using the label information from a single source (sample or center information), which ignored the joint effect of the center and sample information on improving the robustness of pseudo-labeling. To address this issue, we propose Pseudo-labeling Integrating Centers and Samples with Consistent Selection mechanism (PICSCS) for UDA. First, PICSCS assigns the label vector with (1 + K) pseudo labels to the target sample. Specifically, the 1 pseudo label represents the center information, which is determined by its nearest class center. The K pseudo labels symbolize adequate sample information, and they are determined by its K nearest source samples. Second, to use the label information from different sources in pseudo-labeling, PICSCS defines the consistent selection mechanism by judging whether (1 + K) pseudo labels in the label vector are the same. Then, label vectors can be identified as consistent or inconsistent, and only target samples with consistent label vectors are adopted in the iteration. Finally, extensive experiments on four benchmark datasets (ImageCLEF-DA, Office-31, Office-Caltech, and Office-Home) show that PICSCS makes the iteration stable, and PICSCS outperforms the state-of-the-art UDA methods.