An improved generative adversarial network for translating clothes from the human body to tiled image

Abstract

Nowadays, purchasing the products similar to the styles of stars’ products has become a new trend in e-commerce platforms. Clothing transaction constitutes to the major part of these kinds of online purchasing. Traditional methods firstly segment clothes from human body and then input the segmented clothing image patch into a retrieval system as a query in a way that similar clothing items could be retrieved. However, the segmented clothing images usually contain complex backgrounds, and these clothing items appear to be twisted, as they are segmented from human body straightforwardly. In order to assist this cross-scenario clothing retrieval, this paper introduces a new triple-supervised GAN (TripleGAN) model by translating the clothes on human body into tiled clothes. Our model was trained on a large-scale dataset including over 30,000 clothing pairs constructed by ourselves. Extensive experimental results exhibit that our model consistently can generate tiled clothing images with more delicate details and higher quality compared with other models. Our model also shows promising performance in terms of cross-domain clothing retrieval in real-life applications.