Point cloud: what it is, where it comes from, and how to use one in Blender

Blender menu names below are checked against the Blender 5.2 manual and source code.

What is a point cloud?

A point cloud is a sampling of a real surface: thousands or millions of points recorded wherever a measuring device hit an object, a building or a stretch of terrain. At its core, each point is a position, three coordinates X, Y and Z, and nothing links the points together. The Blender manual describes this kind of data as suited to 3D scans, particle simulations and sparse datasets where a surface mesh would be unnecessary or impractical.

That is the key difference from a mesh. A mesh connects its vertices with edges and faces, and faces are what you see when you render it. A point cloud does not describe a surface; it gives you samples of one, denser or sparser depending on the device and the distance.

What each point can store

• Position: X, Y, Z, the only value every point cloud has.
• Color: RGB, taken by the scanner's camera or by the photos of a photogrammetry shoot.
• Intensity: in the LAS specification, the integer representation of the pulse return magnitude, in other words how strong the laser echo was.
• Normal: the direction perpendicular to the surface at that point, which surface reconstruction depends on.
• Classification: in lidar data, a class per point. LAS 1.4 uses 2 for ground, 6 for buildings and 9 for water, among others.

Where point clouds come from

Laser scanners and lidar

NOAA defines lidar (Light Detection and Ranging) as a remote sensing method that uses a pulsed laser to measure ranges. A terrestrial laser scanner on a tripod applies that principle to the scene around a fixed station. Airborne lidar, flown on airplanes and helicopters, pairs every range with the position and orientation of the aircraft from integrated GPS and inertial measurement units, which produces a three-dimensional point cloud of the ground below.

National programs publish this data openly. France's IGN, for example, distributes its LiDAR HD survey, acquired at no less than 10 pulses per square meter (5 above 3,200 m altitude) and classified into 11 categories, as 1 km by 1 km LAZ tiles of roughly 500 MB to 2 GB each, under an open license.

Lidar has also reached consumer hardware. The iPad Pro Apple announced in March 2020 has a LiDAR Scanner that measures the distance to surrounding objects up to 5 meters away, indoors and outdoors.

Photogrammetry

Photogrammetry rebuilds points from many overlapping photos of the same subject. In the AliceVision pipeline behind Meshroom, Structure from Motion solves the position of every camera and outputs a sparse point cloud; a depth map is then computed for each view before meshing and texturing. The guide to turning a photo into a 3D model walks through that workflow.

Depth cameras

A depth camera outputs one distance per pixel. Microsoft's Azure Kinect, for instance, lights the scene with near-infrared light, measures how long it takes to come back (time-of-flight) and stores a Z value in millimeters for every pixel. With the camera's optical parameters, each pixel can be turned into a point.

Point cloud file formats

Blender also reads point clouds from USD files (since 4.2) and glTF files (since 5.2).

What point clouds are used for

• Surveying: keep only the ground class and you get the bare terrain under trees and buildings.
• Architecture and BIM: a scan of an existing building is the base for modeling it.
• Cultural heritage: a monument or a dig is recorded exactly as it stood on a given date.
• Visual effects: a scan of a film set is the reference for lining up CG with the plate.
• Digital twins: a plant or a city is captured so its digital copy can be monitored.

From point cloud to mesh

Texturing, 3D printing or animating a scanned object requires a surface. Getting one usually takes three steps: thin the cloud, compute oriented normals, reconstruct the surface.

Oriented normals first

Poisson reconstruction needs an oriented normal at every input point, which, its authors note, most scanners provide. If your cloud has none, CloudCompare computes them with Edit > Normals > Compute: pick a local surface model (plane, triangulation or quadric), a neighborhood radius and an orientation method such as Minimum Spanning Tree. In MeshLab, the Compute normals for point sets filter starts from 10 neighbors by default.

Poisson surface reconstruction

Michael Kazhdan, Matthew Bolitho and Hugues Hoppe introduced Poisson surface reconstruction at the 2006 Symposium on Geometry Processing. It considers all the points at once, which makes it highly resilient to noise. The screened version, published by Kazhdan and Hoppe in ACM Transactions on Graphics in 2013, creates watertight surfaces from oriented point sets and treats the input points as interpolation constraints, so the result sticks closer to the measurements.

• In MeshLab, the Surface Reconstruction: Screened Poisson filter is driven by Reconstruction Depth, 8 by default: at depth d, the solver grid is no larger than 2^d cells per side.
• In CloudCompare, the qPoissonRecon plugin is driven mainly by octree depth: higher means finer and slower. On open shapes such as terrain, its density field lets you drop the triangles farthest from the input points with Edit > Scalar Fields > Filter by Value.

Either way, the mesh you get is heavy and uneven. The guides to remeshing in Blender and to retopology cover how to turn it into something you can texture and animate.

CloudCompare and MeshLab: the free tools to know

CloudCompare describes itself as 3D point cloud (and triangular mesh) processing software, originally designed to compare two point clouds, such as laser scans. It is free, GPL-licensed and runs on Windows, macOS and Linux. It reads and writes LAS, E57, PLY and ASCII formats, and thins a dense cloud with Edit > Subsample: Random (a set number of random points), Spatial (a minimum distance between points) or Octree (one point per cell).

MeshLab, developed since 2005 by the Visual Computing Lab of ISTI-CNR in Italy under GPL 3.0, cleans, repairs and converts meshes. Its reconstruction filters include Screened Poisson and Ball Pivoting, which builds triangles from the existing points without adding new ones.

Working with point clouds in Blender 5.2

Import a PLY file

1. Go to File > Import > Stanford PLY (.ply), or drag the file into the 3D Viewport.
2. Leave Vertex Colors on sRGB, the default, unless the file stores linear colors (Linear); None skips colors entirely.
3. Leave Merge Vertices off, as it is by default: it merges points that sit at the same location.
4. Check Forward Axis and Up Axis (Y and Z by default) if the cloud comes in lying on its side.

Blender creates a Mesh object named after the file. If the file has no faces, that mesh is vertices only. Colors land in a color attribute called Col, normals in a normal attribute, and any other vertex property, an intensity value for instance, becomes a Float attribute that keeps its name from the file. The viewport draws those vertices as small dots, but a mesh without faces renders as nothing.

The Point Cloud object type

Blender has a dedicated object type for this, Point Cloud. It sat behind an experimental option for a long time and has been fully enabled since Blender 4.5. Cycles has rendered point clouds directly as spheres since 3.1, which the release notes describe as much more efficient in memory and render time than instancing an object on every point. EEVEE draws them too.

• Object > Convert > Point Cloud turns a mesh into a point cloud: one point per vertex, with the mesh attributes, Col included. Object > Convert > Mesh goes the other way.
• Add > Point Cloud creates a cloud of 400 random points, handy for tests.
• Built-in attributes per point are position, radius (its size), color, id and velocity. With no radius attribute, Blender uses 0.01, which is 1 cm at the default unit scale.

Edit Mode has worked on point clouds since 4.5, enough to clean up a scan by hand: select with A, Alt+A, Ctrl+I or Select Random, move points with the transform tools, delete stray points with X, or split an area into its own object with P (Point Cloud > Separate).

Geometry Nodes for points

• Mesh to Points (Mesh > Operations) builds a point cloud from a mesh's vertices, edge midpoints, face centers or corners, with a Radius of 0.05 by default.
• Points to Vertices (Point) does the opposite and outputs one mesh vertex per point.
• Set Point Radius (Point) sets the size of each point, in the viewport and in renders.
• Instance on Points (Instances) places a piece of geometry on every point, a small sphere or a cube for example. Point attributes stay available on the resulting instances.
• Points to Volume (Point) wraps each point in a fog sphere stored in a grid named density. Followed by Volume to Mesh (Volume > Operations), it gives you a mesh without leaving Blender, as fine as the sphere radius and voxel size allow.

Since Blender 4.5, the Geometry Nodes Input > Import menu also has Import PLY and Import CSV. The latter creates a point cloud: every row becomes a point and every numeric column an attribute named after the mandatory header row, with the separator set in Delimiter. Positions stay at zero, so read the x, y and z columns with Named Attribute, combine them with Combine XYZ and feed the result into Set Position.

Show the point colors

1. Select the Point Cloud object and give it a material.
2. In the Shader Editor, add an Attribute node (Shift+A > Input > Attribute) and type the attribute name in Name: Col after a PLY import, or whatever your file calls its color.
3. Connect its Color output to the Base Color of the Principled BSDF, or to an Emission shader to show the recorded colors regardless of scene lighting.
4. Switch the viewport to Material Preview or Rendered to see them.

The Point Info node (Input > Point Info), documented for Cycles, outputs each point's position, radius and a random value, which you can run through a Color Ramp to vary the point colors. The guide to materials in Blender covers the Shader Editor and the Principled BSDF.

A Point Cloud object renders like the rest of the scene; the guide on how to render in Blender covers engine choice and settings. For a flythrough of a dense survey, the frames can be split across several machines with a Blender render farm.

Blender's limits and how to work around them

File > Import in Blender 5.2 has no entry for LAS, LAZ or E57. You have three options.

• Point Cloud I/O (StudioMedio, GPL, on extensions.blender.org) imports and exports E57, PLY, LAS, LAZ, PCD, XYZ and PTS. According to its listing, it creates a native Point Cloud object, keeps colors, normals, intensity and classification as attributes, and its Center on Origin option handles the large coordinates of LAS files. Blender 5.1 or later.
• E57 Point Cloud Importer (Quentin Misslin, GPL, same platform) imports E57 files as Point Cloud objects with colors and normals, from Blender 5.0.1.
• CloudCompare: open the LAS or E57 file, thin it with Edit > Subsample, then save it as PLY with normals, colors and scalar fields.

Add-ons install from Preferences > Get Extensions; the guide to Blender extensions has the details.

Georeferenced coordinates

Lidar surveys often come in map coordinates, around a million meters. Blender's PLY importer stores positions as 32-bit floats, and the CloudCompare wiki, which also works in 32-bit floats, explains that around 10⁶ precision drops to roughly 1 to 10 cm. Move the cloud close to the origin before bringing it into Blender; CloudCompare itself offers a Global Shift beyond about 10⁵.

Gaussian splats, a close cousin

3D Gaussian Splatting, presented at SIGGRAPH 2023 by a team from Inria, Université Côte d'Azur and MPI Informatik (Kerbl, Kopanas, Leimkühler and Drettakis), also starts from a point cloud: the sparse one produced during camera calibration. It then represents the scene with 3D Gaussians, soft oriented blobs, and renders it in real time.