A point cloud is a collection of points in 3D space.
Each point usually has a position:
and may also store extra information like:
- color
- intensity
- surface normal
- timestamp
So point cloud data is one of the simplest ways to represent the shape of a real 3D scene.
Main idea
A point cloud does not explicitly store surfaces, edges, or triangles.
It just stores many sampled points from objects in space.
So if you scan a chair, a road, or a room, the result might be a huge set of 3D points that lie on or near the visible surfaces.
The more points you have, the more detail you usually capture.
Where point clouds come from
Point clouds are often produced by sensors or reconstruction systems such as:
- LiDAR
- depth cameras
- stereo vision
- photogrammetry
For example, a LiDAR sensor sends out light pulses and measures how long they take to bounce back. From that, it estimates distances and builds a 3D set of points.
Why they are useful
Point clouds are useful because they are a fairly direct representation of measured 3D geometry.
They are common in:
- robotics
- self-driving cars
- 3D mapping
- AR/VR
- computer vision
They are often the first usable 3D output you get from a sensor.
Point cloud vs image
An image stores values on a 2D grid of pixels.
A point cloud stores values at irregular 3D locations.
So an image is tied to a camera view, while a point cloud tries to represent the actual geometry of the scene in space.
Point cloud vs mesh
A point cloud only stores points.
A mesh stores:
- vertices
- edges
- faces
So a mesh tells you how points are connected into surfaces, while a point cloud does not.
This means point clouds are simpler, but also less structured.
If you want a solid surface model, you often start with a point cloud and then build a mesh from it.
Why point clouds can feel messy
Point clouds usually have some of these problems:
- noise
- missing regions
- uneven density
- outliers
- no explicit connectivity
That is why point cloud processing often involves:
- filtering
- registration
- segmentation
- surface reconstruction
Intuition
The best mental model is:
A point cloud is a 3D scatter plot of a real scene.
Each point says “something was observed here”.
Enough points together start to reveal the shape of the object or environment.
Connection to other 3D representations
A point cloud is often an intermediate representation.
It can be used to build or support:
- meshes
- occupancy maps
- volumetric models
- scene reconstructions
There is also a connection to 3DGS.
3D Gaussian Splatting does not just use raw points. Instead, it uses many 3D Gaussians with position, size, shape, and appearance. You can think of this as a richer representation than a plain point cloud.
So point clouds are one of the most basic building blocks in 3D data.