Rotation-Invariant Point Convolution With Multiple Equivariant Alignments

TL;DR

This paper introduces a novel rotation-invariant 3D point convolution method that leverages multiple equivariant alignments, significantly improving performance in object classification and segmentation tasks.

Contribution

It demonstrates that rotation-equivariant alignments can be used to create rotation-invariant layers, enhancing 3D deep learning models.

Findings

Achieves state-of-the-art results in 3D object classification.

Reduces performance gap between rotation-invariant and standard models.

Improves semantic segmentation accuracy.

Abstract

Recent attempts at introducing rotation invariance or equivariance in 3D deep learning approaches have shown promising results, but these methods still struggle to reach the performances of standard 3D neural networks. In this work we study the relation between equivariance and invariance in 3D point convolutions. We show that using rotation-equivariant alignments, it is possible to make any convolutional layer rotation-invariant. Furthermore, we improve this simple alignment procedure by using the alignment themselves as features in the convolution, and by combining multiple alignments together. With this core layer, we design rotation-invariant architectures which improve state-of-the-art results in both object classification and semantic segmentation and reduces the gap between rotation-invariant and standard 3D deep learning approaches.