Normal Estimation and Curvature

Treat normal estimation as a throwaway preprocessing call and the damage surfaces much later, where it is hard to trace. Poisson reconstruction grows bubbles and bridges across the surface, half the normals point inward so the shading renders the model inside out, and a registration that should have locked cleanly drifts because the point-to-plane term is constrained by the wrong direction. None of these failures announce themselves at the moment the normals are computed. On a heritage TLS scan, PCA on a 5 cm radius neighbourhood estimates normals on flat ashlar to within 3 to 5 degrees of a hand-fit ground truth, yet at the corners where two stone faces of an arch meet, exactly the points the conservator most needs sharp, the same radius smooths the points into a featureless mean direction with angular errors of 25 to 40 degrees. The fixed radius that gives a clean wall normal gives a useless edge normal. A badly estimated normal fails silently, and the error compounds through every algorithm that consumes it.

Normal estimation is the point in the pipeline where cleaned coordinates begin to acquire geometric meaning, and its inputs are shaped by everything upstream. The sampling density of Chapter 5 determines how many points a local neighbourhood contains, the filtering of Chapter 6 determines how clean those points are, and the spatial indexes of Chapter 4 answer the nearest-neighbour queries that gather them. Downstream, point-to-plane ICP (Chapter 8), region growing by curvature (Chapter 11), the local descriptors of Chapter 13, and screened Poisson reconstruction (Chapter 15) all assume that a reliable oriented normal is already attached to each point. The chapter builds that normal in stages: PCA on a local neighbourhood, the choice of that neighbourhood's size, polynomial jet fitting for curvature, the orientation of the normal's sign, robust and learned variants for edges, outliers, and vegetation, and finally the metrics that reveal whether the estimates can be trusted.

The chapter moves from estimation to orientation to robustness, and we begin with what the normal is for, because the demands of the algorithms that consume it explain the design of the estimators.

The role of surface normals

A raw point cloud is a set of coordinates with no edges, faces, or notion of inside versus outside. The normal defined above is the simplest differential-geometric quantity carried by the surface, and attaching one to every point turns the unstructured set into a geometrically meaningful representation.

Several examples illustrate this. In surface reconstruction (Chapter 15), the Poisson method takes oriented normals as its primary input, and without them it cannot distinguish solid from void. In registration, point-to-plane ICP converges roughly ten times faster than point-to-point ICP, because the normal constrains the optimisation to slide along the surface. In segmentation, region-growing algorithms merge neighbouring points whose normals are nearly parallel and group them into smooth surfaces. In feature description, descriptors such as FPFH and SHOT encode angular histograms relative to the local normal, which makes them rotation-invariant. In rendering, normals determine how light interacts with the surface, and without them a point cloud appears as a flat scatter of dots.