Virtual Shirt Fitting Model based on Deep Learning & Computer Vision

Abstract

This study introduces a novel try-on condition generator, unifying the warping and segmentation generation stages. A newly proposed feature fusion block facilitates information exchange, mitigating misalignment and pixel-squeezing artifacts. Furthermore, the incorporation of discriminator rejection filters ensures the accuracy of segmentation map predictions, enhancing the overall performance of virtual try-on frameworks. Experimental results on a high-resolution dataset showcase the effectiveness of the proposed model in addressing misalignment and occlusion, surpassing baseline models.