Gaussian Splatting

Super Fast Rendering - realtime display rates at 1080p Higher quality than Mip-Nerf Avoids unnecessary computation in empty space

Nerf and voxel-based 3D representations require stochastic sampling for rendering - computationally expensive, result in noise

use sparse point clouds obtained from SFM - no need for MVS 1-5 million gaussians per scene

Rendering: -respect visibility ordering -sorting -tile-based rasterization

opacity-based rendering — —the ‘opacity’ (light absorption) assigned to a point depends on the distance from the previously sampled point, AND the density at the point basically, you are assuming the ‘point’ is a block that absorbs light

These are continuous representations, (not voxels) works with randomly initalized gaussians can project gaussians to 2D and perform alpha blending

Parametrize each gaussian with: opacity, anisotropic covariance, spherical harmonics

adaptive density control - add or remove gaussians during training Last Reviewed: 1/17/25