Reproduce “PoissonNet: A Local-Global Approach for 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
Arman Maesumi, Tanish Makadia, Thibault Groueix, Vladimir G. Kim, Daniel Ritchie, Noam Aigerman
PoissonNet learns mesh features by transforming local surface gradients and using a Poisson solve to produce scalar updates with support across the mesh. In rig-free humanoid reposing, it predicts deformations while retaining detail on tested characters and poses. On MOYO's 27-part segmentation task, it reaches 97.03% test accuracy versus DiffusionNet's 96.12%, specific to that setup.
Key insights
- Poisson's equation, acting as an 'integration' operator, provides an effective mechanism to bridge local gradient information with global scalar signals, enabling a truly global receptive field in each network block.
- By operating in the gradient domain and leveraging Poisson solves, the network can preserve the full frequency spectrum of features, crucial for representing intricate geometric details.
- The use of intrinsic differential operators and efficient sparse linear solvers ensures triangulation agnosticism and high computational efficiency without relying on lossy spectral approximations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.