Accurate 3D Face Reconstruction from Multiple RGBD Images

초록

Reconstructing accurate 3D face models from 2D images remains a key challenge in computer vision. While traditional models like FLAME provide strong facial priors, they struggle with fine details and often lack depth information, leading to imprecise reconstructions. Recent approaches have improved depth accuracy using 3D-2D supervision, but comprehensive pipelines that integrate multi-view inputs and depth refinement are still missing. In this paper, we propose an end-to-end pipeline that refines FLAME's coarse shape using depth data and introduces a novel multi-view texture generation method using Poisson blending to produce high-quality, artifact-free textures.