MM-VTON: A Multi-stage Virtual Try-on Method Using Multiple Image Features

Virtual try-on allows users to see how they look without actually trying the clothes on during their purchase. This technology has numerous applications in the display of clothing effects and is especially useful during the pandemic, because it enables remote try-on without physical contact. The major limitations of current virtual try-on methods, however, lie in the difficulty of addressing clothing deformation, edge synthesis, etc. In this study, we present a new three-stage virtual try-on method to reduce the reliance on clothing regions in human images. To achieve this, we design a new semantic prediction module to fully remove clothing-related information from human images. Additionally, we introduce a new try-on module to fuse the extracted features using an adversarial loss, resulting in significant improvements on the try-on image quality. Experimental results have demonstrated the effectiveness of our method, which achieves competitive results in comparison to state-of-the-art methods.