Exploring Discriminability of Categorical Anchors for Heterogeneous Domain Adaptation*

Heterogeneous Domain Adaptation (HDA) is an attractive task generalizing a model across source domain and target domain with distinct distributions and features. Previous HDA approaches have revealed the effectiveness to obtain shareable features between heterogeneous domains by benefiting from the pseudo labels of target data. Unfortunately, there are two primary limitations in the existing heterogeneous domain adaptation methods. Firstly, target pseudo labels are a two-edged sword to feature alignment. Secondly, The decision boundaries are not adequately discriminative since the margin among different categories of the target domain are not explicitly maximized. In this paper, we propose an HDA method that unifies categorical anchors of the target domain to align the feature distribution between domains and avoid direct utilization of target pseudo labels. Specifically, the intra-class feature variance and the interclass feature variance between source and target domain are simultaneously optimized to learn shared discriminative features. The margin among different categories of the target domain is explicitly maximized to make the decision boundaries clearer. To reduce the bias of limited supervision of target domain, this paper proposes to select the categorical anchors in the shared subspace to boost the performance of classifier since they are reliable to characterize the distribution of different categories. Comprehensive experiments verify that the proposed method can gain remarkable margins over state-of-the-art visual domain adaptation methods.