Reproduce “From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data” 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
Yuming Zhao, Junhui Hou, Qijian Zhang, Jia Qin, Ying He
Zhao et al. introduce PRISM, a 3D pre-training framework that learns intrinsic geometric features by predicting geodesic distances directly from unstructured point clouds. The method achieves a Mean Relative Error of 3.87% in geodesic prediction and enables the first successful feed-forward surface parameterization model, demonstrating strong performance across multiple geometry-sensitive downstream tasks.
Key insights
- Leveraging geodesic distance as an intrinsic metric provides representations invariant to isometric deformations, capturing fundamental manifold structure.
- Learning a latent feature space where Euclidean distances approximate geodesic distances allows for inherently robust geometric features.
- Combining a geodesic structure consistency loss with importance sampling effectively guides topological learning and mitigates prediction bias across distance ranges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.