The normal filter returns the estimated normal and curvature for a collection of points. The algorithm first computes the eigenvalues and eigenvectors of the collection of points, which is comprised of the k-nearest neighbors. The normal is taken as the eigenvector corresponding to the smallest eigenvalue. The curvature is computed as

\[curvature = \frac{\lambda_0}{\lambda_0 + \lambda_1 + \lambda_2}\]

where \(\lambda_i\) are the eigenvalues sorted in ascending order.

The filter produces four new dimensions (NormalX, NormalY, NormalZ, and Curvature), which can be analyzed directly, or consumed by downstream stages for more advanced filtering.

The eigenvalue decomposition is performed using Eigen’s SelfAdjointEigenSolver.

Normals will be automatically flipped towards the viewpoint to be consistent. By default the viewpoint is located at the midpoint of the X and Y extents, and 1000 units above the max Z value. Users can override any of the XYZ coordinates, or set them all to zero to effectively disable the normal flipping.


By default, the Normal filter will invert normals such that they are always pointed “up” (positive Z). If the user provides a viewpoint, normals will instead be inverted such that they are oriented towards the viewpoint, regardless of the always_up flag. To disable all normal flipping, do not provide a viewpoint and set always_up to false.

Default Embedded Stage

This stage is enabled by default


This pipeline demonstrates the calculation of the normal values (along with curvature). The newly created dimensions are written out to BPF for further inspection.



The number of k-nearest neighbors. [Default: 8]
A single WKT or GeoJSON 3D point. Normals will be inverted such that they are all oriented towards the viewpoint.
A flag indicating whether or not normals should be inverted only when the Z component is negative. [Default: true]