Cross-domain Cross-task Knowledge Distillation Network for Unsupervised Domain Adaptation*

Abstract Unsupervised domain adaptation has attracted extensive attention in recent years because it can naturally solve the problem of new domains encountered in depth model testing. Generally, unsupervised domain adaptive methods combine classification loss and adversarial loss, so that the features extracted from the model can be classified and are domain independent, thus alleviating the problems caused by cross domain models. However, it is impossible for the features extracted from the model to not contain any domain information. Therefore, we propose a new solution to make the source domain features extracted from the model fully include the style of the target domain images. Specifically, a compression network is trained by the target domain images, and the source domain features extracted by the classification network are distilled with the target domain image features extracted by the compression network to make it have the style of the target domain images. Finally, our method achieves leading or comparable results on five public data sets.