VoroFacet: Junction-aware Mesh Extraction
from Neural Unsigned Distance Fields

Anonymous authors

Under double-blind review

Single-layer meshes from neural UDFs, with incident sheets joined along shared junctions.

Abstract

Unsigned distance fields (UDFs) can represent surfaces of arbitrary topology, but recovering mesh connectivity from neural UDFs remains challenging, especially at non-manifold junctions where multiple sheets meet. To this end, we propose VoroFacet, a novel pipeline that extracts single-layer meshes from neural UDFs and reconstructs junction connectivity accurately through Voronoi facets. Specifically, we first use a near-surface double-layered proxy derived from the UDF as a local spatial partition for facet selection, where the facets belong to the Voronoi diagram of the sites generated by an adaptive ball cover of the proxy. We then estimate junctions from the mesh of this first extraction and place balls along them, so that in a re-extraction process, the resulting cospherical sites make incident sheets share Voronoi vertices and edges at these junctions. Finally, we propose a three-stage optimization to refine the geometry of the extracted mesh. Experiments demonstrate superior reconstruction quality over state-of-the-art methods, particularly at non-manifold junctions. The source code will be publicly available.

Method overview

An adaptive ball cover of a near-surface double-layered proxy guides Voronoi facet selection. Junction-aware re-extraction makes incident sheets share Voronoi vertices and edges. Remeshing and three-stage optimization refine the extracted geometry.

VoroFacet pipeline: neural UDF, double-layered proxy, adaptive ball cover, initial surface extraction, junction-aware re-extraction, remeshing and geometry refinement.

From a double-layered proxy to a single-layer mesh with shared junctions. Figure 2.

Full resolution ↗

Qualitative comparisons

All-method comparison on a non-manifold object, with magnified junction regions.

Junction connectivity on a non-manifold object.

Full resolution ↗

Interactive comparisons

18%
Section plane Advanced
50%

Fixed X/Y/Z cut; stays in place when you rotate the view.

0% 100%
Interactive mesh
Not loaded
Interactive mesh
Not loaded
Interactive mesh
Not loaded
Interactive mesh
Not loaded
Interactive mesh
Not loaded
Interactive mesh
Not loaded

Choose a model, then load.

Views synchronized

Viewer guide & model info

Navigation

Choose methods
Use each view’s selector
Rotate
Left-drag
Pan
Ctrl + left-drag, or middle/right-drag
Zoom
Shift + left-drag: ↓ in / ↑ out; scroll / pinch / + / −
Set orbit center
Double-click a visible surface
Fit + clear view clip
R or Reset view
Exit large view
Esc

View clipping

Near removes the front; Far removes the back. Both follow your viewing direction.

Near clip
Ctrl + scroll · 1% per step
Far clip
Ctrl + Shift + scroll
Fine adjustment
Add Alt · 0.2% per step
Clear without moving
C or Clear view clip

On Mac, use ⌘ instead of Ctrl. Sliders provide the same control on touch screens.

Fixed section (advanced)

Slice along the model’s X, Y or Z axis. Unlike view clipping, this plane stays fixed when you rotate.

  1. Open Section plane above.
  2. Enable it, choose an axis, then move the slider.
  3. Uncheck Enable section to remove the cut.

Model information

Original model data is cached in this browser for later visits when space permits. Private browsing or clearing site data may remove it.

View 1: not loaded

View 2: not loaded

View 3: not loaded

View 4: not loaded

View 5: not loaded

View 6: not loaded

  • Original meshes; shared coordinates and camera. No simplification or per-method alignment.
  • Mesh edges show triangulation, not junction labels.
  • All clipping is display-only and uncapped.

Quantitative results

65.80%NM-F1
37.07%VA-NM-F1
1.380
Components

Swipe to see all metrics →

Non-manifold Objects · 50 shapes · Table 1
MethodCD ↓× 10³F1 ↑%NM-F1 ↑%VA-NM-F1 ↑%Components

Mean ± std. Best in bold, excluding Components. “—”: no explicit non-manifold edges.

Metric definitions

All methods query the same learned UDF. NM-F1 compares reconstructed and reference junction curves; VA-NM-F1 additionally checks the number of incident sheets. CD and F1 measure surface accuracy.

Ablation studies

Junction-aware re-extraction

Junction-aware re-extraction: all panels from Figure 14 of the paper.
Incident sheets meet along a single shared curve. Figure 14.

Geometry-adaptive remeshing

Geometry-adaptive remeshing: all panels from Figure 15 of the paper.
Sampling follows the local geometry around sharp features. Figure 15.

Three-stage refinement

Three-stage refinement: all panels from Figure 7 of the paper.
From the extracted mesh to the refined surface and junction. Figure 7.
Further comparisons

Resources

Research code will be made publicly available.

Figure preview
Original ↗
100%

Scroll / pinch to zoom · Drag to pan · Esc to close