Optimization of Albedo Map in 3D Face Reconstruction
Mingtong Zhang, Jiaxin Tong, Qiuzi Huang, Yang Liu · 2024
With the advancement of 3D computer vision and image processing techniques, 3D face reconstruction has become an important research area. The two most important parts in the reduction from 2D image to 3D modeling are shape and texture maps. However, due to the limited information that can be obtained from a single image, and the light source in a natural scene will have a large impact on the image texture reduction, so the restored image is mostly distorted. To solve this problem, this paper proposes a gan model for optimizing albedo maps based on position map, which decouples light and shadow and albedo by randomly exchanging texture maps of different photos, and introduces Gaussian noise to improve the robustness of the network and restore more realistic 3D face texture maps. Our network is called Pose2albedo (position map to albedo map), and the experimental results show that the texture images optimized by our model are significantly more effective than the texture maps before optimization in terms of metric benchmarks.