Dual-focus memory contrastive learning for active domain adaptation

Active Domain Adaptation (ADA) aims to significantly enhance model adaptation performance by labeling a small portion of target domain samples. While recent studies have focused exten sively on selecting target domain samples through active sampling strategies, the effective utilization of these selected samples remains underexplored. Most methods concentrate on identifying the most valuable target samples but fail to establish persistent mechanisms to propagate their knowledge throughout the adaptation process and hindering the ability of model to capture the intrinsic structure of the target domain, leading to the underutilization of valuable samples. Our study introduces a novel approach, Dual-Focus Memory Contrastive Learning for Active Domain Adaptation (DumDA), which aims to optimize the use of selected samples. DumDA achieves a more profound utilization of target domain samples by innovatively orchestrating memory-encoded historical features with real-time batch contrast through dual-focus alignment, which enhances the learning and alignment of sample selection. Additionally, DumDA incorporates a hybrid active selection strategy to select reconstructed samples in a class-balanced manner. Experimental results on multiple standard datasets demonstrate that DumDA significantly improves performance in domain adaptation tasks, showcasing its effectiveness and innovation.