Prototype-Guided Continual Adaptation for Class-Incremental Unsupervised Domain Adaptation

This paper studies a new, practical but challenging problem, called Class-Incremental Unsupervised Domain Adaptation (CI-UDA), where the labeled source domain contains all classes, classes in unlabeled target increase sequentially. problem is due to two difficulties. First, and label sets are inconsistent at each time step, which makes it difficult conduct accurate alignment. Second, previous unavailable current resulting forgetting of knowledge. To address this we propose novel Prototype-guided Continual (ProCA) method, consisting solution strategies. 1) Label prototype identification: identify prototypes by detecting shared with cumulative prediction probabilities samples. 2) Prototype-based alignment replay: based on identified prototypes, align both domains enforce model retain With these strategies, ProCA able adapt class-incremental effectively. Extensive experiments demonstrate effectiveness superiority resolving CI-UDA. The @scut.edu.cnsource code available https://github.com/Hongbin98/ProCA.git .