Class-incremental Source-free Domain Adaptation Based on Multi-prototype Replay andAlignment

Traditional source-free domain adaptation usually assumes that all the target domain data is available,but in practice,the target domain data often appears in the form of streams,that is,the classes in the unlabeled target domain will increase sequentially,which undoubtedly brings new challenges.First,in each time step,the label space of the target domain is a subset of the source domain,and blind alignment will cause the performance of the model to deteriorate.Secondly,in the process of learning new classes,it will destroy the previously learned knowledge,resulting in the catastrophic forgetting of the previous knowledge.In order to solve these problems,this paper proposes a method based on multi-prototype replay and alignment(MPRA).In this method,the shared classes in the target domain are detected by cumulative prediction probabilities,the label space inconsistency problem is solved,and the multi-prototype replay is used to deal with catastrophic forgetting and improve the memory ability of the model.Additionally,the method incorporates cross-domain contrastive learning based on multi-prototype and source model weights to align feature distributions and improve model robustness.A large number of experiments show that the proposed method has achieved superior performance on 3 benchmark datasets.