Load a single-tree LiDAR point cloud. The tool slices it horizontally, clusters each slice, fits a circle to the cluster that continues the trunk, and walks upward — giving you the main stem, its diameter at breast height, and its taper.
A widely used leaf–wood separation tool, GBSeparation, locates the trunk base with an algebraic circle fit that accumulates Σx³, Σy³, Σx²y from raw coordinates and then subtracts those sums.
Σx³
Σy³
Σx²y
On a cloud georeferenced in UTM (x≈5×10⁵, y≈2×10⁶) those terms reach ~10²², and the differences lose every significant digit of float64 to catastrophic cancellation. The fit returns a circle of radius ~10⁶ m, the algorithm's graph root lands millions of metres from the tree, and the segmentation collapses to a thin skeleton.
It is not a noise problem, and not a scanner problem. Terrestrial scans usually sit in local coordinates and work fine; the same tree in UTM does not. Subtracting the horizontal centroid before fitting is the whole fix.
Toggle Auto-centre coordinates off with a georeferenced cloud loaded and the panel will show you both fitted radii side by side.
Everything runs client-side in WebGL and a Web Worker. No point ever leaves your machine — there is no upload endpoint.