Reproduce “Neural Gaussian Scale-Space Fields” 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
Felix Mujkanovic, Ntumba Elie Nsampi, Christian Theobalt, Hans-Peter Seidel, Thomas Leimkühler
Neural Gaussian Scale-Space Fields learn continuous, anisotropic Gaussian scale-space representations directly within neural fields. The method enables efficient, single-pass generation of filtered signals from unfiltered training data, demonstrating competitive or superior performance in image and 3D geometry filtering and practical applications like texture anti-aliasing and robust optimization.
Key insights
- Dampening high-frequency Fourier features in combination with Lipschitz-bounded MLPs allows neural networks to approximate Gaussian filtering.
- This method enables learning continuous, anisotropic Gaussian scale-space representations directly from unfiltered data in a self-supervised manner.
- A post-training calibration step establishes a linear relationship between the learned pseudo-covariance and true Gaussian covariance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.