Equivariant Networks for Pixelized Spheres

Mehran Shakerinava, Siamak Ravanbakhsh

Introduction

Spherical data is ubiquitous across scientific disciplines, from global climate patterns to 360-degree "omnidirectional" photography. However, this data does not naturally fit onto the flat, rectangular grids that standard Convolutional Neural Networks (CNNs) require. To bridge this gap, researchers often project the sphere onto the faces of a Platonic solid (such as a cube or an icosahedron) and then "pixelize" those faces into grids.

While this projection allows the use of grid-based operations, it introduces a major challenge: equivariance. In deep learning, equivariance ensures that if you rotate the input (e.g., a climate map), the network’s internal representation rotates accordingly in a predictable way. Previous methods, like Gauge Equivariant CNNs, treated the rotation of each face independently. This is a "loose" approximation because, in reality, rotating a cube as a whole strictly dictates how the pixel grids on its faces must move and rotate.

This paper introduces a novel framework for deep learning on spherical data by formally modeling the interplay between the global symmetries of Platonic solids and the local symmetries of their pixelized faces. By using group theory to link these two levels of symmetry, the authors create highly expressive equivariant networks that achieve state-of-the-art results.

The Pixelization Approach

To approximate a smooth sphere, the authors utilize the geometric properties of Platonic solids. By starting with a base solid and recursively subdividing its faces into finer grids, the resulting structure can be projected outward to form a high-resolution spherical approximation.

Iterative pixelization: starting from a Platonic solid, faces are subdivided and projected to approximate a sphere.

Methods: Bridging Global and Local Symmetries

The authors propose a new family of equivariant linear maps that treat the pixelized sphere as a unified geometric object rather than a collection of independent, disconnected charts.

  1. Symmetry Coupling: Instead of assuming faces can be swapped or rotated arbitrarily, the authors use semi-direct products and block systems from group theory. This ensures that a global rotation of the solid (like a 90° turn of a cube) correctly "rolls" and translates the feature grids on every face simultaneously.
  2. Equivariant Padding: To allow information to flow seamlessly across the edges of the Platonic solid, the authors developed a specialized padding procedure. This ensures that the boundaries between faces respect the global symmetry of the solid.
  3. Parameter Sharing: The core of their network is a set of linear layers where weights are "tied" or shared based on these symmetry groups. They released a library, AutoEquiv, to automate the creation of these complex parameter-sharing matrices.
  4. Orientation Awareness: For tasks where "up" and "down" matter (like 360° indoor photography), the framework can restrict the symmetry group to only include rotations around the vertical axis.
Global rotations of the solid dictate the transformation of local feature grids.

The "Rolling" Effect

A key insight of this work is that rotating a pixelized cube doesn't just move a face to a new position; it also changes the orientation of the grid on that face. The authors demonstrate that a strict hierarchy—where global and local transforms are treated independently—is less expressive than their "interacting" model. By reducing unnecessary symmetry constraints and formally capturing this rolling effect, the model becomes more flexible and learns more effectively.

Key Results

The proposed "Final Model" consistently outperformed previous hierarchical and gauge-equivariant models across multiple domains:

  • Climate Science (HAPPI20 Dataset): On high-resolution climate simulations used to detect extreme weather events (like tropical cyclones), the model achieved a Mean Accuracy of 99.30% and a Mean Average Precision (mAP) of 95.20%. This is a massive leap over the previous state-of-the-art, which had an mAP of 80.6%.
  • Omnidirectional Vision (Stanford 2D3DS): On a dataset of 1,413 omnidirectional (360°) RGBD indoor images used for semantic segmentation, the "oriented" version of their cube-based model reached 62.5% Mean Accuracy, setting a new benchmark record.
  • Efficiency: By using a "Pool-Broadcast" implementation for their global operations, the authors kept the computational overhead manageable despite the complex symmetry constraints.

Limitations and Future Work

While highly effective, the framework has some limitations. The authors excluded the dodecahedron from their study because its face tiling (using triangles) lacks translational symmetry, making standard convolutions difficult. Additionally, projecting a flat grid onto a sphere inherently introduces non-uniform distortion. While treated as negligible at high resolutions, it remains a theoretical approximation.

Looking forward, the use of "block systems" is a powerful, generalizable tool. The authors suggest this framework could be extended to other complex geometric structures beyond spheres, such as 3D meshes or more exotic manifolds where local and global symmetries are deeply intertwined.

Summary updated

Abstract

Pixelizations of Platonic solids such as the cube and icosahedron have been widely used to represent spherical data, from climate records to Cosmic Microwave Background maps. Platonic solids have w...
BibTeX
@inproceedings{shakerinava2021equivariant,
  author = {Mehran Shakerinava and Siamak Ravanbakhsh},
  title = {{Equivariant Networks for Pixelized Spheres}},
  booktitle = {International Conference on Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=M_TiE6CslH0R},
}
Figure