Register Point Clouds (ICP)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is a single downstream-resolved choice fixed at deploy time: the metrics record, a bare rigid transform, or the transformed source PointCloud. The transformed cloud uses the full un-downsampled source and carries source colours forward while dropping normals.
• Hard constraintBoth inputs must be point arrays shaped as N-by-3 rows; colours and normals, when present, must have the same row count as their points. An empty or degenerate input yields an identity transform with zero fitness rather than a failure.
• Parameter interactionThe colored variant of method requires colours on BOTH clouds and rejects the message otherwise; the point-to-plane variant uses target normals and estimates them from local neighbourhoods when absent, needing at least three target points.
• AccuracyIterative closest point only refines an already-coarse alignment and has no global registration step. Set max_correspondence_distance on the order of the expected initial misalignment, since correspondences beyond it are ignored.
• Parameter interactionvoxel_size downsamples both clouds before matching for speed and robustness; the reported transform aligns the downsampled clouds but is applied to the full source when the cloud arm is emitted. Zero disables downsampling.
• LatencyConvergence stops early once the per-correspondence error plateaus or the incremental translation collapses, so max_iterations is an upper bound rather than a fixed cost.