Domain Adaptation Using the Replay Buffer: Adaptive Sampling Using Domain-Specific Classifier

Domain adaptation is a method used to reduce discrepancy between source and target domains and to enhance generalization performance by transforming the distribution of source images. This study proposes a method to improve domain adaptation, which is based on selecting feature-rich images from their respective domains. Training images are represented based on their proximity to the domain by applying a domain-specific classifier. The method then selects more representative images from the training set using adaptive sampling. Based on the confidence from the classifier, less confident images in the mini-batch are replaced by more confident images in the replay buffer. This approach enhances the quality of the training set, ensuring the model focuses on high-quality, domain-relevant data. Experimental results demonstrate the efficiency of the proposed method, achieving consistent improvements with an average FID reduction of 6.60% across various tasks. Additionally, Additionally, it is extensively shown that the proposed method can improve recent methods on various domain adaptation tasks, both quantitatively and qualitatively.