Reproduce “NeuroGF: A Neural Representation for Fast Geodesic Distance and Path Queries” 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
Qijian Zhang, Junhui Hou, Yohanes Yudhi Adikusuma, Wenping Wang, Ying He
NeuroGF encodes the geodesic relationships of a 3D mesh in a compact neural implicit representation that predicts both point-to-point distances and shortest paths from a single forward pass. Shape-specific models achieve about 0.49% mean relative distance error, path Chamfer-L1 error of 1.25×10^-2, sub-millisecond GPU query times, and roughly 10 MB memory usage, while generalizable variants extend the approach to unseen shapes with higher error.
Key insights
- Offline neural fitting can replace repeated online geodesic computation with compact, GPU-parallel inference, reducing query cost without storing large mesh-dependent geodesic structures.
- Jointly modeling distances, paths, and the surface helps: removing the SDF, path, distance, or geometric consistency components degrades the corresponding predictions.
- Mesh connectivity improves cross-shape prediction, but the overfitted representation remains more accurate than the generalizable variants and does not provide exactness guarantees.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.