VKSR scales kernel surface reconstruction to point clouds with millions of points without requiring subdivision schemes and while maintaining strong performance on sparse inputs, unlike popular Screened Poisson Surface Reconstruction (SPSR).
We propose Vecchia Kernel Surface Reconstruction (VKSR), an accurate implicit surface reconstruction method that efficiently scales recent kernel-based techniques to large point clouds with millions of points.
While existing (global) kernel methods work well in a sparse setting, due to low-rank approximations, performance degrades quickly when presented with dense point clouds sampled from surfaces with high geometric complexity or large-scale inputs with millions of points. To overcome this limitation and inspired by the Gaussian Process literature, VKSR uses Vecchia's approximation instead of low-rank approximations, which naturally shifts computation from a global to a local level and allows reconstructing 14M+ points in minutes.
VKSR achieves state-of-the-art results on several challenging datasets while retaining kernel methods' favorable properties when reconstructing sparse inputs.
The key idea of VKSR is simple: We view kernel surface reconstruction from the Gaussian Process (GP) perspective and replace previously employed low-rank approximations with Vecchia's approximation, routed in the GP literature.
Meshes are extracted at a voxel resolution of 128³ (to ease visualization, we display meshes decimated by a factor of two, with negligible loss of geometric detail). See paper for details.
Drag to rotate, right-drag to translate, scroll to zoom, and press R to reset all views.
Meshes are extracted at a voxel resolution of 256³ (to ease visualization, we display meshes decimated by a factor of ten, with negligible loss of geometric detail). See paper for details.
Drag to rotate, right-drag to translate, scroll to zoom, and press R to reset all views.
Meshes are extracted at a voxel resolution of 1024³ (to ease visualization, we display meshes decimated by a factor of ten, with negligible loss of geometric detail). See paper for details.
Drag to rotate, right-drag to translate, scroll to zoom, and press R to reset all views.
Meshes are extracted at a voxel resolution of 512³ (to ease visualization, we display meshes decimated by a factor of ten, with negligible loss of geometric detail). See paper for details.
Drag to rotate, right-drag to translate, scroll to zoom, and press R to reset all views.
@inproceedings{weiherer2026vksr,
author = {Weiherer, Maximilian and Umah, Chukwudi Williams and Egger, Bernhard},
title = {VKSR: Scalable Kernel Surface Reconstruction Using Vecchia's Approximation},
booktitle = {European Conference on Computer Vision},
year = {2026},
}