Abstract:

The goal of domain adaptation is transferring knowledge learned from the labeled source domain to the unlabeled target domain, where the distributions of data from the source domain and target domain are different. Prior domain adaptation methods typically assume data from the source domain are available when learning to adapt to the target domain, which might be not possible due to privacy issues and data security. In order to transfer knowledge to the target domain without accessing data from the source domain, a new source-free domain adaptation method called FCDC is developed in this paper. SFDC splits data from the target domain into two parts according to their confidence and estimates pseudo labels with different strategies for the two parts. For data of high confidence, their pseudo labels are the predictions of the neural network. For data of low confidence, their pseudo labels are guided by data of high confidence, which is modeled as an optimization problem, and the solution of the problem gives the pseudo label of data of low confidence. To make the pseudo labels of data of low confidence more reliable, FCDC utilizes information maximum loss on these data to produce well-behavior clusters. Meanwhile, FCDC takes advantage of self-supervision loss on data of high confidence to make the features of these data more diverse and surround data of low confidence. Experimental results show that the proposed FCDC is an effective method for source-free domain adaptation.

Key words: source free domain adaptation, pseudo-label, confidence