Adversarial Domain Adaptation Enhanced via Self-training

Deep learning models trained on large number of labeled samples improve the accuracy of many tasks of computer vision. In addition to this, since collecting and labeling vast amount of samples in various domains is difficult, it is important to develop adaptable models to different domains. In unsupervised domain adaptation, given data of labeled samples on source domain, our goal is to learn a classifier which performs well for both the samples on source domain and unlabeled samples on target domain. Although recent adversarial domain adaptation methods made impressive progress, training the classifier on source samples hinders the classifier from perfectly generalizing to the target samples. To this end, we propose an adversarial domain adaptation method enhanced via self-training to overcome the generalization problems of adversarial domain adaptation methods. In order to perform self-training, pseudo labels are assigned to the samples on target domain to learn more generalized representations for target domain. The experimental results on benchmark domain adaptation dataset, VisDA-2017, show that our proposed method significantly improves and outperforms the base method exploited in this work.