Point Cloud Filtering & Voxel Downsampling

Preprocessing steps that crop irrelevant regions, remove noisy points, and shrink a raw point cloud down to a usable size.

Raw point clouds straight from a depth camera or lidar often have hundreds of thousands of points, mixed with stray outlier points from measurement error, and with irrelevant regions like the tabletop or the floor. Feeding this directly into downstream algorithms is both slow and unstable, so it's preprocessed first. Common steps include: cropping by a bounding box to keep only the work region; voxel downsampling, which divides space into fixed-size small cubes (voxels) and merges the points within each one into a single point (PCL takes the centroid, Open3D also averages color and normal), drastically reducing the point count and making density more uniform; and outlier removal, where a statistical method removes points whose average distance to their k nearest neighbors deviates too far from the global mean, and a radius-based method removes points with too few neighbors within a given radius. Both PCL and Open3D provide ready-made functions for these steps.

ExamplePCL's tutorial downsamples one scan with a 1 cm voxel, reducing the point count from 460,400 to 41,049. 3D Diffusion Policy (DP3) instead first crops out the tabletop and background with a bounding box, then keeps only 512 or 1,024 points via farthest point sampling.

Also called

Voxel Downsampling, Voxel Grid Filter, Outlier Removal

Related

Point Cloud · Voxel · Farthest Point Sampling · Point Cloud Registration · Flying Pixels · 3D Diffusion Policy

Sources

Open3D: Point cloud outlier removal
PCL Tutorial: Downsampling a PointCloud using a VoxelGrid filter
3D Diffusion Policy (arXiv 2403.03954)