---
title: Point Cloud Voxel Downsampling
emoji: 🧊
colorFrom: blue
colorTo: gray
sdk: static
pinned: false
---

# Voxel downsampling with geometry preservation

How far `bigpointcloud_001.ply` (11,357 points, normals + RGB) can be voxel-downsampled before its
geometry degrades, measured rather than eyeballed: surface displacement (mean / RMS / p95 /
Hausdorff), normal deviation and bounding-box shrinkage across a sweep of voxel sizes, plus a
comparison against random subsampling at the same point budget.

- **safe** — voxel 0.210, 8,832 pts (78%), mean error 0.032, normals within 5.7°
- **aggressive** — voxel 0.279, 5,592 pts (49%), mean error 0.083, normals within 13.3°

At the same 5,592-point budget, random subsampling matches the mean error but is 3× worse at the
Hausdorff bound. See `index.html` for the full write-up, `voxel_downsample.py` for the script.