Kim Junho ; Kim Minjae ; Kang Hyeonwoo ; Lee Kwanghee

Abstract

We propose a novel method for unsupervised image-to-image translation, whichincorporates a new attention module and a new learnable normalization functionin an end-to-end manner. The attention module guides our model to focus on moreimportant regions distinguishing between source and target domains based on theattention map obtained by the auxiliary classifier. Unlike previousattention-based method which cannot handle the geometric changes betweendomains, our model can translate both images requiring holistic changes andimages requiring large shape changes. Moreover, our new AdaLIN (AdaptiveLayer-Instance Normalization) function helps our attention-guided model toflexibly control the amount of change in shape and texture by learnedparameters depending on datasets. Experimental results show the superiority ofthe proposed method compared to the existing state-of-the-art models with afixed network architecture and hyper-parameters. Our code and datasets areavailable at https://github.com/taki0112/UGATIT orhttps://github.com/znxlwm/UGATIT-pytorch.