Author: Kerbl
Year: 2024
- represents a scen with 3D gaussians
- Init the process with camera poses and points from SfM
- (1-5 Millions of Gaussians per scene)
- Traditionnal method = reproject and blend Multiview stereo (MVS) from geometry
- starts with COLMAP, creates a set of 3D gaussians with the pc (mean, covariance and opacity
$\alpha$ ), colors are represented with spherical harmonics (SH) (anisotropic color representation) - project 3D gaussians in 2D for rendering
- to keep properties of 3D covariances (spd) parametrize them as an ellipsoid (scaling vector s + rotation matrix)
- optimize both gaussians (
$p, \Sigma, \alpha, SH$ ) and the guassian density (adaptive control of gaussian density) - optimization based on both rendering and comparing
- the first gaussian from Sfm has uniform colors
- remove transparent gaussian (with
$\alpha$ below a thresh) - heuristics to handle the density of gaussians
- differentiable tile based rasterizer
- Memory consumption higher than NerFs (~20Go st)