Domain Adaptation via Image to Image Translation

Unsupervised Domain Adaptation (UDA) has recently attracted a lot of attention from the computer vision community. In this chapter, we review a general framework for UDA, which allows deep neural networks trained on a source domain to be tested on a different target domain without requiring any training annotations in the target domain. UDA is a challenging problem, as it aims at overcoming the potentially large difference between source and target data distributions, which is known as the “domain gap.” Here we propose a general UDA algorithm by adding extra networks and losses that help regularize the features extracted by the backbone encoder network. To this end we propose the novel use of the recently proposed unpaired image-to-image translation framework to constrain the features extracted by the encoder network. We leverage three main ideas: (1) we require that the features extracted by encoders are able to reconstruct the images in both domains, hence, encoders provide pseudo-invertible nonlinear mappings, (2) we require that the distribution of features extracted from images in the two domains to be indistinguishable in the encoders’ output space (i.e., the latent space), (3) we require various cycle consistencies on source and target encoders and decoders. Many recent works can be seen as specific cases of our general framework. We apply our method for domain adaptation between MNIST, USPS, and SVHN datasets, and Amazon, Webcam and DSLR Office datasets in classification tasks, and also between GTA5 and Cityscapes datasets for a segmentation task. We demonstrate state of the art performance on each of these datasets.