NDNet: natural deformation of apparel for better virtual try-on experience

ABSTRACTVirtual try-on networks is a fascinating application of generative models in fashion, providing users the ease of trying on the apparel of their choice in the comfort of their homes. Several challenges have to be addressed by these networks to provide an experience that emulates a physical changing room, which includes feature preservation, natural deformation of apparel, realistic generation of occluding body parts, etc. In this work, we present Natural Deformation Network (NDNet), which aims for a more natural deformation of apparel with respect to the pose of the person while conserving the important details of the apparel as well as the person. We propose a novel Siamese Attention UNet generator to fulfill this objective. By leveraging the benefits of the Siamese architecture and attention gates in the generator, our work shows a commendable improvement in both quantitative as well as qualitative performance compared to the widely-acclaimed CP-VTON, as well as significant qualitative improvements over other works. The Siamese Attention UNet generator can be employed in future works as well for the task of naturally deforming the apparel.