MACA: Multi-Anchor Classification Approach for Unsupervised Domain Adaptation

Abstract

Unsupervised Domain Adaptation for image classification aims to adapt models trained on a labeled source domain to an unlabeled target domain, improving target domain classification performance. However, previous UDA classification researches tend to assume the two domain distributions after domain adaptation are the same, ignoring their differences. Such unrealistic assumptions may reduce the effectiveness of domain adaptation. To tackle this challenge, we propose the Multi-Anchor Classification Approach for Unsupervised Domain Adaptation (MACA) to model two domain data distributions simultaneously. Specifically, we create a Gaussian distribution for each class to characterize the general data distribution for each class, and sample from it to obtain multiple sub-centers. These sub-centers represent different patterns of data variation between two domains, thus yielding a more robust representation of the features and facilitating auxiliary classification. We conducted extensive experiments on three widely adopted UDA benchmarks. The results demonstrate that MACA significantly outperforms current domain adaptation methods.