Active domain adaptation with mining diverse knowledge: An updated class consensus dictionary approach.

Domain adaptation (DA) has recently emerged as an effective paradigm for training the target model with labeled source knowledge. When knowledge transfer in DA encounters the bottleneck, one effective crack way is to introduce the labeled data to guide the DA process. Along this line, many active learning-based DA approaches are emerging to improve the quality of samples selection at the decision border. However, these methods have not preserved and exploited cross-domain common knowledge. In this work, we propose active domain adaptation with mining diverse knowledge: an updated class consensus dictionary approach (UCCDA). Specifically, we firstly initialize the class consensus dictionary by the source prior knowledge. Then, we choose high-confident target pseudo samples through self-training, while assigning labels to those low-confident via oracle annotation. In addition, we design the class consensus dictionary to guide the alignment between the source and target domains, instead of traditional direct cross-domain data alignment. Remarkably, to prevent error accumulation during the consensus dictionary learning, we specially design the anti-forgetting mask matrix to randomly restore the original knowledge. Finally, abundant experiments demonstrate that UCCDA outperforms the related state-of-the-art approaches. [ABSTRACT FROM AUTHOR]