A Generative Adversarial Distribution Matching Framework for Visual Domain Adaptation

Abstract

One of the difficulties in computer vision is how to build an accurate classifier for a new target domain with insufficient labeled images from a related source domain with labeled images. Adversarial learning is a novel domain adaptation method that tackles this challenge by training robust deep networks and reducing the distribution difference between source and target domains, thus improving the classification performance on a target task. However, most prior adversarial adaptation learning approaches merely reduce the distribution difference across domains through GAN (Generative Adversarial Networks)-based loss, but when the performance of a generator or discriminator in GAN is degraded, the distribution difference between source and target domains are difficult to decrease. In this paper, we propose a novel generalized framework for adversarial domain adaptation, referred to as Generative Adversarial Distribution Matching. Our idea is to add the data discrepancy distance between source and target domains to the objective function of the generator so as to reduce distribution difference across domains through a Generator and a Discriminator compete against each other. Comprehensive experimental results confirm that it can well outperform several state-of-the-art methods for cross-domain image classification problems.