Shape-preserving Swapping Autoencoder for Garment Texture Transfer

Abstract

Garment image synthesis became possible since GAN has achieved great success in the field of image generation. However, the works in this area are limited. This paper is focused on solving the problem of swapping the texture of clothing images. Existing methods like SwappingAE and style transfer can not solve this problem well. A shape-preserving swapping autoencoder (SP-SwappingAE) is proposed to solve the clothing swapping problem. Comparing to SwappingAE, we proposed a condition discriminator to retain the structure of input images. To verify our proposed method, we collect a clothing dataset, named FCI, including 60,000 different types of upper garment images. Experiment results on FCI dataset showed that our proposed method beat SwappingAE and the style transfer algorithm. The high-resolution result shows that the swapped images have realistic texture and structure. In future works, we will explore more clothing design methods with artificial intelligence.