High-Level Semantic Feature Matters Few-Shot Unsupervised Domain Adaptation

In few-shot unsupervised domain adaptation (FS-UDA), most existing methods followed the learning (FSL) to leverage low-level local features (learned from conventional convolutional models, e.g., ResNet) for classification. However, goal of FS-UDA and FSL are relevant yet distinct, since aims classify samples in target rather than source domain. We found that insufficient FS-UDA, which could introduce noise or bias against classification, not be used effectively align domains. To address above issues, we aim refine more discriminative Thus, propose a novel task-specific semantic feature method (TSECS) FS-UDA. TSECS learns high-level image-to-class similarity measurement. Based on features, design cross-domain self-training strategy few labeled build classifier addition, minimize KL divergence distributions between domains shorten distance two Extensive experiments DomainNet show proposed significantly outperforms SOTA by large margin (i.e., ~10%).