Skip to content

Poisson Reconstruction

In one sentence: Reconstruct a watertight closed mesh from a point cloud with per-point normals using the Poisson surface reconstruction algorithm.

Overview

Workbench: Reverse Engineering
Menu: Reverse Engineering → Surface Reconstruction → Poisson…
Shortcut: none

Solves a Poisson equation over the space around the point cloud to find an implicit surface whose gradient matches the input normals, then extracts the zero-level set as a triangle mesh. The result is always watertight (closed, no holes) regardless of scan gaps in the input.

Intuition

Imagine the surface you want to reconstruct as an invisible soap bubble. You have a cloud of points on that bubble, each with a little arrow (normal) pointing outward. Poisson reconstruction treats those arrows as constraints and solves a global equation to find the bubble shape that best fits all the arrows at once. Because it solves globally, it naturally fills in gaps in the scan — the bubble stays closed even where you have no data.

The trade-off is that it always closes the surface. If you scanned only one side of an object, Poisson will invent a smooth cap on the unseen side.

When to use it

  • You have a dense, well-oriented point cloud with normals and want a clean, closed mesh for further modelling.
  • Your scan has small gaps (missed regions) that you want filled automatically.
  • You need a watertight solid for FEM simulation or 3-D printing from a reconstructed scan.

When NOT to use it

  • When your cloud has no normals — Poisson requires per-point outward normals. A plain XYZ cloud will fail or produce garbage. Estimate normals in CloudCompare first.
  • For open surfaces (one side of an object) — Poisson will add an invented cap. Use mesh segmentation + B-spline approximation instead.
  • For sharp features — Poisson smooths everything. Thin walls, sharp creases, and fine details are rounded. Use the Screened Poisson variant (available in CloudCompare) for better feature preservation.

Step-by-step

  1. Import a point cloud with per-point normals into FreeCAD (the normals must be pre-computed externally, e.g. in CloudCompare).
  2. Select the Points::Feature object.
  3. Choose Reverse Engineering → Surface Reconstruction → Poisson….
  4. In the task panel, set:
  5. Octree depth — start with 8; increase for finer detail.
  6. Samples per node — start with 1.0; increase for smoother output.
  7. Click OK. FreeCAD creates a Mesh::Feature of the reconstructed mesh.

Tip

Octree depth 8 works for most mechanical parts. For detailed organic shapes or fine surface textures, try depth 10. Above depth 12 the computation time grows exponentially.

Parameters

Parameter Type Default Description
Octree depth Integer 8 Controls reconstruction resolution. Higher = more detail, longer computation. Typical range: 6–10.
Solver divide Integer 6 Depth at which the solver subdivides. Typically octree depth − 2.
Samples per node Float 1.0 Minimum sample points per octree cell. Higher = smoother but less detailed.
Scale Float 1.1 Ratio of reconstruction cube size to point bounding box size.

Common mistakes and pitfalls

Reconstruction fails or produces a mess — missing normals

Cause: The point cloud has no per-point normals, or the normals are inconsistently oriented (some pointing inward, some outward).
Fix: Compute and orient normals in CloudCompare before importing (Edit → Normals → Compute, then Edit → Normals → Orient).

Reconstructed mesh adds unwanted caps on open scans

Cause: Poisson always creates a closed surface. Open scan data gets closed with an invented smooth cap.
Fix: Either trim the cap manually with the Mesh workbench, or use B-spline approximation (Approximate B-Spline Surface) for open surfaces.

Fine details are smoothed away

Cause: Poisson is a global smoothing solver. Sharp edges and thin walls are inherently smoothed.
Fix: Use higher octree depth and lower samples per node. For very sharp features, post-process in Meshmixer or use a different algorithm.

Python API

import ReverseEngineering as Reen
import FreeCAD as App

doc = App.ActiveDocument
pts = doc.getObject("PointsWithNormals")   # Points::Feature with normals

mesh_feat = doc.addObject("Mesh::Feature", "PoissonMesh")
# TODO: verify exact Reen API for Poisson
mesh_feat.Mesh = Reen.poissonReconstruction(
    Points=pts.Points,
    OctreeDepth=8,
    SolverDivide=6,
    SamplesPerNode=1.0
)
doc.recompute()

See also