Self Separation and Misseparation Impact Minimization for Open-Set Domain Adaptation

Most of the existing domain adaptation algorithms assume the label space in the source domain and the target domain are exactly the same. However, such a strict assumption is difficult to satisfy. In this paper, we focus on Open Set Domain adaptation (OSDA), where the target data contains unknown classes which do not exist in the source domain. We concluded two main challenges in OSDA: (i) Separation: Accurately separating the target domain into a known domain and an unknown domain. (ii) Distribution Matching: deploying appropriate domain adaptation between the source domain and the target known domain. However, existing separation methods highly rely on the similarity of the source domain and the target domain and have ignored that the distribution information of the target domain could help up with better separation. In this paper, we propose a algorithm which explores the distribution information of the target domain to improve separation accuracy. Further, we also consider the possible misseparated samples in the distribution matching step. By maximizing the discrepancy between the target known domain and the target unknown domain, we could further reduce the impact of misseparation in distribution matching. Experiments on several benchmark datasets show our algorithm outperforms state-of-the-art methods.