Separation of Unknown Features and Samples for Unbiased Source-free Open Set Domain Adaptation

Abstract

Open Set Domain Adaptation (OSDA) is proposed to train a model on a source domain that performs well on a target domain with domain discrepancy and unknown class samples outside the source domain. Recently, Source-free Open Set Domain Adaptation (SF-OSDA) aims to achieve OSDA without accessing source domain samples. Existing SF-OSDA only focuses on the known class samples in the target domain and overlooks the abundant unknown class semantics in the target domain. To address these issues, in this paper, we propose a Separation of Unknown Features and Samples (SUFS) method for unbiased SF-OSDA. Specifically, SUFS consists of a Sample Feature Separation (SFS) module that separates the private features from the known features in each sample. This module not only utilizes the semantic information of each sample label, but also explores the potential unknown information of each sample. Then, we integrate a Feature Correlation Representation (FCR) module, which computes the similarity between each sample and its neighboring samples to correct semantic bias and build instance-level decision boundaries. A large number of experiments in the SF-OSDA scenario have demonstrated the effectiveness of SUFS. In addition, SUFS also shows great performance in the Source-free Partial Domain Adaptation (SF-PDA) scenario.