Image to Image Translation With Multi- Scale Generator

Abstract

Image to image translation is a highly generalized learning task, that canbe applied to a wide number of Computer Vision application domains. Condi-tional Generative Adversarial Networks (cGANs) are used to perform image toimage translation. The generator network typically used in the existing cGANapproach, Pix2Pix, adopts the U-Net architecture, consisting of encoding anddecoding convolutional layers and skip-connections between layers of the sameresolution. While effective and convenient, such an arrangement is also restrictivein some ways, as the feature reconstruction process in the decoder cannot utilizemulti-scale features. In our work, we study a generator architecture where featuremaps are propagated to the decoder from different resolution levels. We’ve exper-imentally shown improved performance on two different datasets — the NYU-V2depth dataset and the Labels2Facades dataset.