Open-Set Domain Adaptation Under Few Source-Domain Labeled Samples

Sayan Rakshit, Balasubramanian S, Hmrishav Bandyopadhyay, Piyush Bharambe, Sai Nandan Desetti, Biplab Banerjee, Subhasis Chaudhuri; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 4029-4038

Recently, the notion of closed-set few-shot domain adaptation (FSDA) has been introduced where limited supervision is present in the source domain. However, FSDA overlooks the fact that the unlabeled target domain may contain new classes unseen in the source domain. To this end, we introduce the novel problem definition of few-shot open-set DA (FosDA) where the source domain contains few labeled samples together with a large pool of unlabeled data, and the target domain consists of test samples from the known as well as new categories. We propose an end-to-end model called FosDANet to tackle such a scenario which operates on two principles: to generate confident pseudo-labels for the unlabeled source samples and to classwise align the source and target domains for the known classes while rejecting the unknown-class data. A combination of a self-supervised loss and a novel triplet-based relation learning module is devised to aid in confident pseudo-labeling, and a dual adversarial learning scheme is proposed for domain alignment. Extensive experiments were performed on five datasets: Office-31, Office-Home, Adoptiape, and two new datasets we designed: mini-domainnet and a remote sensing benchmark called NPU-RSDA. FosDANet is found to consistently outperform the relevant literature.