Bi-directional class-wise adversaries for unsupervised domain adaptation

摘要

Unsupervised domain adaptation relies on well-labeled auxiliary source domain information to get better performance on the unlabeled target domain. It has shown tremendous importance for various classification and segmentation problems. Classical methods rely on diminishing the domain discrepancy in the latent space but ignore class-wise information, which will lead to elimination of the inherent data structure. To avoid destroying the inherent structure during unsupervised domain adaptation, we propose a Bi-Directional Class-level Adversaries cross-domain model (BDCA) with two symmetric classifiers interpolating two latent spaces to build a tunnel between the source domain and target domain. Specifically, we propose a class-level discrepancy metric to enforce domain consistency during the trend of domain adaption. We also employ two symmetric classifiers that are collectively optimized to maximize the discrepancy on target sample prediction. Extensive experiments are conducted on four publicly available datasets (i.e. office-31, office-home, GTAV and Cityscapes) and two challenging computer vision prediction problems, i.e., image classification and semantic segmentation. Quantitative and qualitative results demonstrate the effectiveness of our proposed model.