Textured Mesh Generation Using Multi-View and Multi-Source Supervision and Generative Adversarial Networks

Abstract

This study focuses on reconstructing accurate meshes with high-resolution textures from single images. The reconstruction process involves two networks: a mesh-reconstruction network and a texture-reconstruction network. The mesh-reconstruction network estimates a deformation map, which is used to deform a template mesh to the shape of the target object in the input image, and a low-resolution texture. We propose reconstructing a mesh with a high-resolution texture by enhancing the low-resolution texture through use of the super-resolution method. The architecture of the texture-reconstruction network is like that of a generative adversarial network comprising a generator and a discriminator. During the training of the texture-reconstruction network, the discriminator must focus on learning high-quality texture predictions and to ignore the difference between the generated mesh and the actual mesh. To achieve this objective, we used meshes reconstructed using the mesh-reconstruction network and textures generated through inverse rendering to generate pseudo-ground-truth images. We conducted experiments using the 3D-Future dataset, and the results prove that our proposed approach can be used to generate improved three-dimensional (3D) textured meshes compared to existing methods, both quantitatively and qualitatively. Additionally, through our proposed approach, the texture of the output image is significantly improved.