Reproduce “3D-Rotation-Equivariant Quaternion Neural Networks” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Wen Shen, Binbin Zhang, Shikun Huang, Zhihua Wei, Quanshi Zhang

A framework for 3D-Rotation-Equivariant Quaternion Neural Networks (REQNNs) systematically revises existing network architectures to imbue them with robust 3D rotation equivariance and permutation invariance for point cloud processing. This approach achieves superior classification accuracy on arbitrarily rotated data, reaching 83.03% on ModelNet40, and enables direct manipulation of feature orientation for reconstruction tasks.

Key insights

  • By systematically modifying layer-wise operations and removing bias terms for convolution, networks can intrinsically maintain rotation equivariance when processing quaternion features.
  • The norm of a quaternion provides a rotation-invariant scalar, which can be leveraged for rotation-invariant outputs (e.g., classification) while maintaining equivariance throughout feature extraction.
  • Principal Component Analysis (PCA) can be used to transform 3D coordinates into a local reference frame that rotates with the input, enabling equivariant processing of coordinate-dependent operations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.