Can You Convert Images Between Cats And Humans? Transgaga", A Development Of CycleGAN

3 main points
✔️ Image style transformation task 
✔️ Separate input images into geometry (orientation) and appearance (shape) 
✔️ Enables style transformations that require large geometric changes, which is not possible with CycleGAN 

TransGaGa: Geometry-Aware Unsupervised Image-to-Image Translation
written by Wayne Wu, Kaidi Cao, Cheng Li, Chen Qian, Chen Change Loy
(Submitted on 21 Apr 2019)
Comments: Accepted to CVPR 2019. Project page: this https URL
Subjects: Computer Vision and Pattern Recognition (cs.CV)
 

Introduction 

Although pix2pix was proposed as a style transformable model, pix2pix required the preparation of paired datasets. However, unlike pix2pix, CycleGAN introduced cycle loss to enable style transformation without paired datasets. However, style transformations requiring large geometric changes did not work well, and only color changes between horses and zebras were possible.

MUNIT, proposed in 2018, used VAE (Variational AutoEncoder) to provide a common variable between the two domains. The goal of representing two domains with one latent variable is intuitive, but in practice, it was not so easy and difficult to represent the geometric structure between more complex domains.

The model proposed in this study divides the image into geometry and appearance, and reduces them to latent variables, enabling style transformations that require large geometric changes. In addition, we obtained more natural and diverse generated images than conventional methods.

Y.Tanabe