Abstract
In computer vision, when labeled images of the target domain are highly insufficient, it is challenging to build an accurate classifier. Domain adaptation stands for an effective solution to address it by utilizing available and related source domain which has sufficient labeled images, even when there is a substantial difference in properties and distributions of these two domains. Yet, most prior approaches merely reduce subspace conditional or marginal distribution differences between domains but entirely ignoring label dependence (LD) information of source data in subspace. This article proposes a novel approach of domain adaptation, called enhanced subspace distribution matching (ESDM), which makes good use of label information to enhance the distribution matching between the source and target domains in a shared subspace. It reduces both conditional and marginal distributions in a shared subspace during a procedure of kernel principal dimensionality reduction and also preserves source data LD information to the maximum extent, thereby significantly improving cross domain subspace distribution matching. We also provide a learning algorithm with highly affordable computation, which solves the ESDM optimization problem without using time-consuming iterations. Results confirm that it can well outperform several recent domain adaptation methods on image classification tasks in terms of classification accuracy and running time. The results can be used in social cognition, person reidentification, and human-machine interactions.