Stochastic Poisson Surface Reconstruction with One Solve using Geometric Gaussian Processes

Authors: Sidhanth Holalkere, David Bindel, Silvia Sellán, Alexander Terenin

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Results show that our approach provides a cleaner, more-principled, and more-flexible stochastic surface reconstruction pipeline. 4. Experiments and Applications We now demonstrate the proposed approach empirically on a suite of example problems.
Researcher Affiliation Academia 1Cornell University 2Columbia University 3MIT. Correspondence to: Sidhanth Holalkere <EMAIL>.
Pseudocode No The paper describes methods in prose and mathematical formulations, but does not contain a dedicated 'Pseudocode' or 'Algorithm' section or block.
Open Source Code Yes Code: HTTPS://GITHUB.COM/SHOLALKERE/GEOSPSR.
Open Datasets Yes Meshes. The Armadillo, Bunny, Falcon, Scorpion, Springer, Tree, and Well meshes are from Oded Stein’s repository, at: ODEDSTEIN.COM/MESHES. The Dragon mesh is originally from the Stanford 3D Scanning Repository we use the version from Alec Jacobson’s repository, at: GITHUB.COM/ALECJACOBSON/COMMON-3D-TEST-MODELS/.
Dataset Splits No The paper discusses input data and test points but does not specify explicit training/test/validation dataset splits, percentages, or methodologies for data partitioning.
Hardware Specification Yes All reported timings are calculated on a machine running Ubuntu 20.04 with an Intel Xeon Silver 4316 CPU, 256GB RAM, and an Nvidia RTX A6000 GPU.
Software Dependencies No We implement our algorithm in Python using GPYTOOLBOX [29] for common geometry processing subroutines, JAX [7] for numerical computations, and render our results in Blender using BLENDERTOOLBOX [21].
Experiment Setup Yes All of our results use the Matérn kernel with ν = 3/2 and length scales between 4 × 10^−2 and 1 × 10^−2, depending on the specific mesh. Unless specified otherwise, we use L = 100^3 functions for the cross-covariance and L˜ = 40^3 functions for the prior samples. The cross-covariance is amortized using a total of 50^3 points.