Learning Category Discriminability for Active Domain Adaptation

Active Domain Adaptation (ADA) attempts to improve the adaptation performance on a target domain by annotating informative target data with a limited budget. Previous ADA methods have significantly advanced by incorporating domain representativeness and predictive uncertainty. However, they only focus on domain-level alignment and ignore the category discriminability of two domains, which may cause classwise mismatching. These mismatched data are overlooked by the above query strategy. To solve this, a Learning Category Discriminability approach is proposed for active domain adaptation. Specifically, it achieves semantic-level alignment and selects the informative target data consistent with the domain adaptation based on task-specific classifiers. To overcome the class imbalance from the small queried data, progressive augmentation of the queried set with confident pseudo labels is designed in our work. In addition, discriminability and diversity learning for unlabeled target samples are performed to improve the reliability of pseudo labels, which makes the classification boundaries more applicable to the target domain. Extensive experiments on two benchmarks, Office-31 and Office-Home, demonstrate the superiority of the proposed method.