Reproduce “MetaSDF: Meta-learning Signed Distance Functions” 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
Vincent Sitzmann, Eric R. Chan, Richard Tucker, Noah Snavely, Gordon Wetzstein
Researchers from Stanford University and Google Research developed MetaSDF, a meta-learning framework for learning 3D shape spaces using Signed Distance Functions. This method enables rapid adaptation of a meta-initialized neural network to represent new shapes with few gradient steps, achieving reconstruction quality comparable to auto-decoder methods at significantly faster inference speeds and demonstrating improved generalization.
Key insights
- Framing shape space learning as a meta-learning problem allows for rapid adaptation of a meta-initialized neural network to represent new, specific shapes.
- The meta-initialization embeds an implicit prior over the entire shape distribution, enabling fast specialization with minimal observations.
- MetaSDF combines the inference speed of encoder-decoder methods with the flexibility and generalization capabilities of auto-decoder frameworks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.