Cross-Domain Contrastive Learning for Unsupervised Domain Adaptation

Unsupervised domain adaptation (UDA) aims to transfer knowledge learned from a fully-labeled source different unlabeled target domain. Most existing UDA methods learn domain-invariant feature representations by minimizing distances across domains. In this work, we build upon contrastive self-supervised learning align features so as reduce the discrepancy between training and testing sets. Exploring same set of categories shared both domains, introduce simple yet effective framework CDCL, for alignment. particular, given an anchor image one domain, minimize its cross-domain samples class relative those categories. Since labels are unavailable, use clustering-based approach with carefully initialized centers produce pseudo labels. addition, demonstrate that CDCL is general can be adapted data-free setting, where data unavailable during training, minimal modification. We conduct experiments on two widely used benchmarks, i.e., Office-31 VisDA-2017, classification tasks, achieves state-of-the-art performance datasets.