CanonNet: Spectral Canonicalization and Curvature-Driven Learning for Compact Local-Geometry Point-Cloud Operators

Benjy Friedmann, Michael Werman

Keywords: Geometric deep learning, Invariant models, Canonicalization

Abstract:

To address the persistent challenges of scalability and robust local geometry representation in point-cloud processing, we propose CanonNet, a highly efficient local feature operator.
CanonNet first employs spectral canonicalization to establish an invariant local frame for each neighborhood. It then uses a geometric learning framework, trained on synthetic surfaces, to distill fundamental curvature priors into a lightweight MLP.
This design allows CanonNet to achieve competitive performance on various benchmarks with approximately 100X fewer parameters, while also exhibiting robust domain transfer.
Its efficiency and design make it an effective building block for deep, hierarchical models, acting as a geometric analogue to the convolution operator for capturing multi-scale features.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 11790