However, the disclosure is not limited thereto and the basis point set may be sampled from a rectangular grid, a ball grid, hexagonal close packing (HCP), and/or other shapes using any technique known to one of skill in the art without departing from the disclosure. Each basis point may be associated with a particular position represented using a three-dimensional coordinate system.

In some examples, the basis point set may be selected from the random uniform ball without further processing, which may be referred to as an unordered basis point set. However, the disclosure is not limited thereto and to improve a performance of a trained model, the system 100 may optionally process the unordered basis point set to introduce a notion of neighborhood or locality, which may be referred to as an ordered basis point set. For example, the system 100 may order the basis points using a k-D tree, such as arranging the points in a k-D tree and sorting them according to leaf indices. As local computations in neural networks over spatially correlated basis points in the feature set benefit from this ordering (e.g., using convolutions or locally connected layers), processing the point cloud data using the ordered basis point set improves a performance of the system 100.