High Resolution UDF Meshing via Iterative Networks

TL;DR

This work tackles the challenge of high-resolution meshing for neural unsigned distance fields (UDFs), which inherently produce noise and holes due to the absence of sign changes. It introduces an iterative, neighborhood-aware neural refinement process that progressively propagates information from neighboring voxels and previously extracted surfaces to produce more accurate and complete meshes. By conditioning per-cell predictions on both local UDF/gradient data and the evolving surface context across iterations, the approach achieves state-of-the-art results at high resolutions on diverse shapes, outperforming single-pass baselines and many existing UDF meshing methods. The method offers robust high-resolution surface extraction with practical improvements in accuracy and surface completeness, while noting tradeoffs in speed and the potential need for post-processing in complex scenes.

Abstract

Unsigned Distance Fields (UDFs) are a natural implicit representation for open surfaces but, unlike Signed Distance Fields (SDFs), are challenging to triangulate into explicit meshes. This is especially true at high resolutions where neural UDFs exhibit higher noise levels, which makes it hard to capture fine details. Most current techniques perform within single voxels without reference to their neighborhood, resulting in missing surface and holes where the UDF is ambiguous or noisy. We show that this can be remedied by performing several passes and by reasoning on previously extracted surface elements to incorporate neighborhood information. Our key contribution is an iterative neural network that does this and progressively improves surface recovery within each voxel by spatially propagating information from increasingly distant neighbors. Unlike single-pass methods, our approach integrates newly detected surfaces, distance values, and gradients across multiple iterations, effectively correcting errors and stabilizing extraction in challenging regions. Experiments on diverse 3D models demonstrate that our method produces significantly more accurate and complete meshes than existing approaches, particularly for complex geometries, enabling UDF surface extraction at higher resolutions where traditional methods fail.

Figures (15)

• Figure 1: Avoiding holes in a plane cockpit. (Top) Compared to the ground-truth UDF of a plane, an approximate neural UDF may fail to reach zero, shown in white, at the true surface location. This is clear in the UDF detail shown in the upper right corner, where the two white areas are disconnected. This yields disconnections and holes in the reconstructions, especially at high resolutions. (Bottom) This causes holes in the surfaces reconstructed by current methods. By leveraging neighboring surfaces, our iterative method properly recovers the cockpit nevertheless.
• Figure 2: Iterative Network. Our model iteratively refines cell configurations. The input consists of the UDF values and gradients at the vertices of the target cell, as well as the current estimated sign configurations of the target and neighboring cells. The surface within a given cell, shown in green, is progressively improved by enforcing consistency with its neighbors.
• Figure 3: Qualitative comparison with existing methods (auto-decoders). Surface meshing results of neural UDFs with all methods at resolution of 512 (and 256 for MGN). UNDC failed at resolution 512 due to high GPU memory requirements. $^\dagger$ indicates that the method is combined with DMUDF.
• Figure 4: Qualitative comparison with existing methods (reconstruction from point clouds). Surface meshing results of neural UDFs at resolution of 512. Top: CAP-L 3D scenes. Middle: CAP-L car. Bottom: DiffUDF cars. $^\dagger$ indicates that the method is combined with DMUDF.
• Figure 5: Meshing at different resolutions. While NSD-UDF Stella24 retrieves most of the surface well at a low resolution, it struggles at higher ones. In contrast, our method, recovers the surface well at all resolutions. We use Marching Cubes with both methods.