3D Grid Based Virtual Trial Room

Abstract

Image based virtual trial room technologies are used for integrating modern in-store clothes into a person image which have caught the interest of research as well as representatives of the multimedia and computer vision communities. However, it’s indeed challenging. However, most existing image-based virtual trial room techniques combine both person and in-store clothes images without taking into consideration of mutual relation. An ideal process will not only change the target clothing into the best suitable shape but it will maintain the cloth uniqueness in the resulting image like color, shade, logos and texture of the material that represent the primary clothes. Prior Generative Adversarial Network (GAN) approaches failed to achieve the above essential performance requirements for realistic virtual trial room performance because they do-not manage considerable spatial misalignment between the primary image and targeted cloth. We present a novel fully-learn-able 3D grid virtual trial room for overcoming all significant barriers in this project. In the first stage, it performs an affine transformation and then thin plate spline transformation for matching the in-store clothes to the target person’s body shape using the Geometry Matching Component. As a result, the warped clothes in the shop appear more realistic. We use Transformation Guided Component that generates a composition mask to blend the warped garments and the image produced is guarantee smoothness, which reduces borderline distortions of warped clothes and creates its outcomes more realistic. Numerous trials on the fashion data-set show that our model delivers decent performance in both qualitative and quantitative terms.