Evaluated Distribution Alignment for Heterogeneous Image Recognition

Heterogeneous domain adaptation has achieved promising results in solving the problem that source domain and target domain are sampled from not only different distributions but also different types of features. However, existing methods often only conduct distributed matching or feature alignment due to the challenges of HDA. Additionally, most methods often treated the marginal distribution adaptation and conditional distribution adaptation equally, which would result in poor performance in real applications. In this paper, we propose a Evaluated Distribution Alignment (EDA) approach to address these challenges. Our method is powered by two tactics: feature alignment and distribution alignment. It adaptively leverages the importance of the marginal and conditional distribution discrepancies and several transfer learning algorithms can be regarded as special cases of it. To evaluate the proposed method, we conduct extensive experiments on image classification, which demonstrate the effectiveness of our proposed algorithms over several state-of-the-art methods.