Semi-Supervised Image-to-Image Translation using Latent Space Mapping

Abstract

Labels to Street Scene Face to Cartoon input output Map to Aerial Profile to Frontal Face input output input output Edges to Shoes input output input output Labels to Facade input output Figure 1: We introduce a novel framework for semi-supervised image translation which can be applied to various image-to-image translation tasks. The key idea is to apply the transformation over the image features instead of the original images, resulting in improved quality of translation results even using a small number of paired data. Here we show results of the proposed method on several examples.