[Title] HUGS: Human Gaussian Splats
[Keyword] 3D Gaussian Splatting, Novel-pose Synthesis, Novel-view Synthesis
[Journal] CVPR, 2024
[arXiv] https://arxiv.org/abs/2311.17910
[Summary]
Human Gaussian Splats(HUGS) is a novel method that uses Gaussian Splatting for a human. Unlike a parametric method that uses SMPL parameters, the method using Gaussian Splatting can express more natural meshes without any constraint on the number of vertex. First, HUGS initializes a template mesh called the Da-Pose (a.k.a. the Vitruvian pose), situated in a canonical space where it remains unaffected by translation or rotation. After that, HUGS finds the edge of the template mesh and divides it in half, which increases the number of mesh vertex. As a result, HUGS uses 110,220 vertices and 220,416 faces. It is projected into the space called a feature triplane, that predicts the parameter of 3D Gaussians and LBS weights. Finally, it deforms 3D Gaussians by using predicted LBS weights and pre-obtained SMPL params.
댓글남기기