Definition
A point cloud is a collection of 3D points sampled from a surface. It is simple and lightweight, but it lacks connectivity: no explicit surface is defined between the points.
Point clouds are what many sensors (LiDAR, depth cameras) and reconstruction methods (Structure from Motion, Multi-View Stereo) produce. To get a surface, the points have to be converted, for example with Delaunay triangulation or Poisson surface reconstruction. 3D Gaussian Splatting can be seen as a point cloud where every point is a Gaussian.
Appears in
- Lecture 1, 3D representations: one of the boundary representations for shapes
- Lecture 1, Structure from Motion: SfM outputs a sparse point cloud
- Lecture 1, Gaussian Splatting: each point carries a Gaussian
- Lecture 5, Point Clouds: pros and cons
- Lecture 5, Poisson: oriented point cloud → mesh
- Lecture 5, Procrustes and ICP: aligning point clouds
- Lecture 6.2, Recap: unordered sets, uneven density; learning on point clouds with PointNet, PointNet++, Point Transformer
- Lecture 8, Point-based rendering: surface and volume splatting, 3DGS
- Lecture 9, Point maps: point cloud vs. point map