LAUSR

Domain-to-domain translation methods map images from a source domain to corresponding images from a target domain. The two domains contain images from the same classes, but these images look different. Recent approaches use generative adversarial networks in various configurations and architectures to perform the translation. By using GANs, they inevitably inherit their problems like training instability and mode collapse. We propose a novel approach to the problem that does not use a GAN. Instead, it relies on an hierarchical architecture that encapsulates information of the target domain by using individually trained networks. This hierarchical architecture is then trained as one unified deep network. Using this approach, we show that images from the original domain are translated to the target domain both for the case when there is a one-to-one correspondence between the images of the two domains and for the case that such correspondence information is absent. We visualize and evaluate the translation from one information domain to the other and discuss the proposed model’s relation to the conditional generative adversarial networks. We further argue that deep learning can benefit from the proposed hierarchical architecture.