Learning Domain Adaptive Features with Unlabeled Domain Bridges
This addresses a practical scenario in cross-domain adaptation where source and target domains are largely gapped, which is incremental over conventional methods.
The paper tackles the problem of large domain discrepancy in unsupervised domain adaptation by proposing a novel approach using unlabeled domain bridges, achieving effective image-to-image translation and feature alignment between distantly distributed domains.
Conventional cross-domain image-to-image translation or unsupervised domain adaptation methods assume that the source domain and target domain are closely related. This neglects a practical scenario where the domain discrepancy between the source and target is excessively large. In this paper, we propose a novel approach to learn domain adaptive features between the largely-gapped source and target domains with unlabeled domain bridges. Firstly, we introduce the framework of Cycle-consistency Flow Generative Adversarial Networks (CFGAN) that utilizes domain bridges to perform image-to-image translation between two distantly distributed domains. Secondly, we propose the Prototypical Adversarial Domain Adaptation (PADA) model which utilizes unlabeled bridge domains to align feature distribution between source and target with a large discrepancy. Extensive quantitative and qualitative experiments are conducted to demonstrate the effectiveness of our proposed models.