Skip to main content

IPPR 004

· One min read

Point Cloud

  • The simplest form of a 3D model, a collection of individual points plotted in 3D space.
  • Photogrammetry: science of making measurements from photographs
    • it uses photos of an object taking a different locations
  • Vertices -> Edges -> Faces -> Polygons -> Surfaces
  • PCL: Point Cloud Library

NeRFs

Gaussian Splatting

  • Instead of building objects using polygons, it represents everything using millions of tiny soft 3D shapes called Gaussians.
  • Less computational costs, more efficient.
  • Run in parallel using GPU rasterization.
  1. First, it builds a rough point cloud from images
  2. Then, replaces those points with these Gaussian blobs
  3. It will optimize them until it match original photos as closely as possible.