Generalized Source-free Domain Adaptation

• DOI: 10.48550/arxiv.2108.01614
• Publication Date: 2021-01-01
• Authors: Shiqi Yang; Yaxing Wang; Joost van de Weijer; Luis Herranz; Shangling Jui
• Keywords: Domain Adaptation; Multi-source

Domain adaptation (DA) aims to transfer the knowledge learned from a source domain an unlabeled target domain. Some recent works tackle source-free (SFDA) where only pre-trained model is available for However, those methods do not consider keeping performance which of high practical value in real world applications. In this paper, we propose new paradigm called Generalized Source-free Adaptation (G-SFDA), needs perform well on both and domains, with access current data during adaptation. First, local structure clustering (LSC), aiming cluster features its semantically similar neighbors, successfully adapts absence data. Second, sparse attention (SDA), it produces binary specific activate different feature channels meanwhile will be utilized regularize gradient keep information. experiments, our method par or better than existing DA SFDA methods, specifically achieves state-of-the-art (85.4%) VisDA, all domains after adapting single multiple domains. Code https://github.com/Albert0147/G-SFDA.