Universal Domain Adaptation through Self Supervision

• DOI: 10.48550/arxiv.2002.07953
• Publication Date: 2020-01-01
• Authors: Kuniaki Saito; Donghyun Kim; Stan Sclaroff; Kate Saenko
• Keywords: Discriminative model; Domain Adaptation

Unsupervised domain adaptation methods traditionally assume that all source categories are present in the target domain. In practice, little may be known about category overlap between two domains. While some address settings with either partial or open-set categories, they particular setting is a priori. We propose more universally applicable framework can handle arbitrary shift, called Domain Adaptative Neighborhood Clustering via Entropy optimization (DANCE). DANCE combines novel ideas: First, as we cannot fully rely on to learn features discriminative for target, neighborhood clustering technique structure of self-supervised way. Second, use entropy-based feature alignment and rejection align source, reject them unknown based their entropy. show through extensive experiments outperforms baselines across open-set, open-partial settings. Implementation available at https://github.com/VisionLearningGroup/DANCE.