LWSFACE: Light-Weighted Weakly-Supervised 3D Facial Texture Reconstruction From a Single Image
Shushan Qu, Mingtong Zhang, Jiaxin Tong, Ruochen Zhang, Yang Liu, Yunpeng Jia, Jie Li, Ningwei Xie · 2023
In recent years, digital human generation based on 3D face reconstruction has found extensive utility across various do-mains. The methods of 3D face reconstruction based on the 3D Morphable Model have achieved promising results. However, the fidelity of the reconstructed facial texture is limited due to the low-dimensional representation of 3DMM. Recently, some texture refinement works are proposed to ad-dress this problem, but they requires either high-resolution 3D dataset or complicated network structure. In this paper, we propose LWSFace, which trains a light-weighted network to recover high-fidelity facial texture from 3DMM coeffi-cients and texture extracted from image via estimated UV map in weakly-supervised manner. The results of both qual-itative and quantitative comparison experiment indicate that our LWSFace surpasses state-of-the-art approaches.