CascadeGAN: A category-supervised cascading generative adversarial network for clothes translation from the human body to tiled images.

With the popularity and development of the Internet, purchasing items that are similar to those that appear in videos or on fashion websites has gradually become an established trend in online commerce. Currently, determining how to accurately locate these similar items in a huge e-commerce database poses a major challenge. When processing images that contain clothes worn on the human body, traditional methods usually detect the clothes in these images first, add them to a database for clothing retrieval, and finally identify a few pieces of clothing with a high degree of similarity for recommendation. Generative adversarial networks (GANs) have been widely and successfully utilized in the field of image-to-image translation. In this paper, GAN is used to transfer human body images into tiled clothing images, which can be directly used for clothing retrieval. To achieve this, a category-supervised GAN under a cascading structure is proposed. For model training, a large-scale dataset was compiled that contains 39,521 image pairs. Experimental results demonstrate that the tiled clothing images generated by our proposed method deliver higher quality, as well as performance superiority, for clothing retrieval in comparison to other existing methods. [ABSTRACT FROM AUTHOR]