Three-dimensional point clouds are unordered sets of points, and their orientation is arbitrary. This rotational and permutation ambiguity makes it hard for standard machine learning architectures to process them directly, because the same object can be represented in infinitely many ways.
LexiCan, a new method described on arXiv, addresses this by canonicalizing the point cloud. It selects a unique orbit representative—a single canonical version of the cloud—under rotations and permutations. This representative retains the full geometric information of the original cloud, so no data is lost in the process.
Because the canonical form is unique, standard architectures such as Transformers can be applied directly to point clouds, without the need for specialized equivariant layers. The approach is presented as a general preprocessing step that could simplify model development for 3D data.