Reliable Cluster-Based Framework for Open Set Domain Adaptation

Abstract

Existing Domain Adaptation (DA) has been able to transfer knowledge from a labeled source domain to an unlabeled target domain nicely. However, in a more complex Open Set Domain Adaptation (OSDA) setting containing unknown target domain categories, previous DA methods fail or even negatively transfer. Recently, various OSDA methods have been proposed and achieved good results. Among them, cluster-based methods such as Domain Consensus Clustering have a higher upper boundary because they mine the intrinsic structure of the target domain instead of treating all the unknown classes of the target domain as one category. However, the impact of faulty pseudo-labels suppresses the performance. Instead, we propose a Reliable Cluster-based Framework (RCF), including Coarse Target Clustering, Structured Matching Strategy, and Reliable Pseudo-Label Training modules, as a general framework to solve the impact of faulty pseudo-labels for cluster-based OSDA methods. Experiments on extensive benchmarks demonstrate that RCF significantly outperforms previous state-of-the-arts.