Collaborative learning-based unknown-class instance identification for open-set domain adaptation.

For domain adaptation in open-set scenarios, target domain samples may be collected from unknown object categories, which are not associated with the original source domain. It is important to judge if a target instance is from one of the classes shared by the source and target domains. Toward this end, we propose a Collaborative learning-based unknown-Class Instance Identification (CoCII) model, in which a cross-domain network and a dedicated network are jointly optimized. The knowledge is learnt from the labeled source data, and then leveraged to predict the labels of target domain samples by the first network. The second network specializes in the target domain under the guidance of the first one. We further incorporate an augmented classification head together with semantic-based contrastive regularization into the dedicated network. This will enable the model in capturing information useful for identifying unknown-class instances, as well as identifying the shared classes. The dedicated network in turn guides the cross-domain network via consistency regularization. Empirical results on Office-31/Home, DIGIT and VisDA-2017 demonstrate that CoCII can outperform other existing state-of-the-art approaches in terms of the average of class-wise accuracies over both known and unknown classes. [ABSTRACT FROM AUTHOR]