Virtual Garment Fitting Through Parsing and Context-Aware Generative Adversarial Networks with Discriminator Group

• Main Authors: Su, Wei-Hong, Chen, Sze-Ann, Chin, Chen-I, Hsiao, Hsu-Feng
• Format: Conference Proceeding
• Language: English
• Subjects: Image synthesis; Industries; Morphology; Semantics; Shape; Torso
• ISSN: 2640-0103
• DOI: 10.1109/APSIPAASC58517.2023.10317305

Owing to the rapid growth of the e-commerce industry, image-based virtual try-on has emerged as a popular research topic in recent years. Despite the introduction of multiple approaches to achieve this concept, there remains ample scope for research and improvement. In this regard, Generative Adversarial Networks (GANs) demonstrate a framework possessing immense potential for subsequent development. Nonetheless, the generated images reported in the literature often manifest blurred edges between semantic regions, thereby diminishing the credibility of results. Furthermore, the generation of try-on images may retain the original shape of the upper body clothing on the model mistakenly, such as the length and tightness of the torso, rather than adapting to the shape of the target clothing. In this paper, we propose a more comprehensive architecture to overcome the limitations of GAN-based approaches, which includes the following contributions. First, we introduce a new parsing and context generator that takes into account the warped binary mask of the geometric matching image of the target clothing. The outputs of this generator incorporate the generation of human parsing images that correspond to the generated try-on images. Moreover, we have designed a novel discriminator group that is specifically focused on judging whether the generated image is a reasonable representation of the specific clothing being worn. According to the experimental results, our method effectively exhibits better synthesis quality and remedies the common challenges encountered while using GANs for virtual try-on.