Reproduce “Learning Steerable Filters for Rotation Equivariant CNNs” 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
Maurice Weiler, Fred A. Hamprecht, Martin Storath
Steerable Filter CNNs (SFCNNs) are introduced, an architecture that achieves joint equivariance under translations and rotations by design, addressing a core limitation of conventional CNNs. The framework improves generalization and reduces sample complexity, achieving a 0.714% test error on rotated MNIST and top-tier performance on the ISBI 2012 2D EM segmentation challenge.
Key insights
- Parametrizing filters with circular harmonics allows for analytically exact rotation through simple phase manipulation, avoiding common interpolation artifacts.
- Encoding continuous rotation equivariance directly into the network architecture significantly improves generalization to unseen rotations and reduces training data requirements.
- Adapting weight initialization for composite filters is crucial for the stable training of deep networks that leverage such steerable filter parametrizations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.