Reproduce “DiffusionNet: Discretization Agnostic Learning on Surfaces” 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
Nicholas Sharp, Souhaib Attaiki, Keenan Crane, Maks Ovsjanikov
DiffusionNet learns surface features that can transfer across different mesh discretizations and point clouds. It replaces mesh-specific convolution and pooling with learned diffusion times and local gradient features, using geometric operators for each input representation. On isotropically remeshed FAUST test shapes, it achieves correspondence error 0.68, versus 35.29 for ACSCNN; point-cloud performance worsens, with point-based baselines competitive.
Key insights
- A simple combination of learned diffusion and spatial gradient features can achieve state-of-the-art results for deep learning on surfaces.
- Learning a per-channel diffusion time allows the network to automatically optimize the spatial support of information propagation.
- Grounding the network in fundamental differential geometry operators (Laplacian and gradient) leads to inherent robustness across varying discretizations and representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.