30 Jun 2026
Earlier this month, our PhD student Julius Hege travelled to Denver to present our latest work at the 3D4S Workshop on 3D Geometry Generation for Scientific Computing, held at CVPR, one of the world's leading computer vision conferences.
Earlier this month, our PhD student Julius Hege travelled to Denver to present our latest work at the 3D4S Workshop on 3D Geometry Generation for Scientific Computing, held at CVPR — one of the world's leading computer vision conferences.
The Problem
Reconstructing accurate 3D shapes from point clouds — scattered, sparse, or noisy sets of 3D coordinates — is a classic challenge in computer vision. Think of it as scanning a real-world object with a lidar sensor: you get millions of points, but turning them into a clean, watertight 3D surface is hard, especially when the data is sparse or noisy.
The Method
nPSR is a hybrid method that combines a well-established mathematical technique (Poisson Surface Reconstruction) with a modern neural architecture. Concretely, we solve the classical Poisson equation but replace the classical solver with a Fourier Neural Operator. This lets the network learn a data-driven prior while staying grounded in principled mathematics to reliably generalize to new data.
Key Result
The model trains on low-resolution grids but generalizes to higher resolutions without any retraining. This "resolution-agnostic" performance can be attained efficiently.
Read the paper here
Authors: Hector Andrade Loarca, Julius Hege, Daniel Cremers, and Gitta Kutyniok