Unsupervised Multi-Domain Image Translation with Domain-Specific Encoders/Decoders

摘要

Unsupervised Image-to-Image Translation achieves spectacularly advanceddevelopments nowadays. However, recent approaches mainly focus on one modelwith two domains, which may face heavy burdens with large cost of $O(n^2)$training time and model parameters, under such a requirement that $n$ domainsare freely transferred to each other in a general setting. To address thisproblem, we propose a novel and unified framework named Domain-Bank, whichconsists of a global shared auto-encoder and $n$ domain-specificencoders/decoders, assuming that a universal shared-latent sapce can beprojected. Thus, we yield $O(n)$ complexity in model parameters along with ahuge reduction of the time budgets. Besides the high efficiency, we show thecomparable (or even better) image translation results over state-of-the-arts onvarious challenging unsupervised image translation tasks, including face imagetranslation, fashion-clothes translation and painting style translation. Wealso apply the proposed framework to domain adaptation and achievestate-of-the-art performance on digit benchmark datasets. Further, thanks tothe explicit representation of the domain-specific decoders as well as theuniversal shared-latent space, it also enables us to conduct incrementallearning to add a new domain encoder/decoder. Linear combination of differentdomains' representations is also obtained by fusing the corresponding decoders.