RealDA: Bridging the domain gap with realistic synthesized images for unsupervised domain adaptation

Abstract

To alleviate the gap between source and target domains in the unsupervised domain adaptation (UDA), existing studies primarily focus on generating high-quality pseudo-labels for unlabeled target samples. However, when pseudo-label quality is not guaranteed, performance can degrade significantly. In this paper, we present a simple yet effective approach, RealDA. Since annotations of the target domain are not accessible, RealDA leverages a well-pretrained Stable Diffusion model to synthesize target-style images from text prompts, allowing their actual ground-truth labels to be used for training and thereby ensuring label quality. Furthermore, to diversify the training data, we introduce the SynPaste strategy, which mixes synthesized and real target images with a soft-label generation scheme. Experiments on three widely used UDA benchmarks (Office-Home, Office-31, and DomainNet) demonstrate the effectiveness of RealDA.