Cross-Domain Interpolation for Unpaired Image-to-Image Translation

Abstract

Unpaired Image-to-image translation is a brand new challenging problem that consists of latent vectors extracting and matching from a source domain A and a target domain B. Both latent spaces are matched and interpolated by a directed correspondence function F for \(A \rightarrow B\) and G for \(B \rightarrow A\). The current efforts point to Generative Adversarial Networks (GANs) based models due they synthesize new quite realistic samples across different domains by learning critical features from their latent spaces. Nonetheless, domain exploration is not explicit supervision; thereby most GANs based models do not achieve to learn the key features. In consequence, the correspondence function overfits and fails in reverse or loses translation quality. In this paper, we propose a guided learning model through manifold bi-directional translation loops between the source and the target domains considering the Wasserstein distance between their probability distributions. The bi-directional translation is CycleGAN-based but considering the latent space Z as an intermediate domain which guides the learning process and reduces the inducted error from loops. We show experimental results in several public datasets including Cityscapes, Horse2zebra, and Monet2photo at the EECS-Berkeley webpage (http://people.eecs.berkeley.edu/~taesung_park/CycleGAN/datasets/). Our results are competitive to the state-of-the-art regarding visual quality, stability, and other baseline metrics.