Meshing Non-Watertight Surfaces from Unsigned Distance Fields

This direction is merged into another direction: watertight manifold constraint. Its last version stays readable and citable here.

Version 1 of 114 papers21 conceptsfrozen 2026-10-04

Traditional 3D reconstruction often relies on Signed Distance Fields (SDFs), which represent a surface as the boundary between "inside" (negative distance) and "outside" (positive distance). While powerful, this representation enforces a "watertight" constraint: every object must be a closed volume. This fails for real-world geometries like clothing, leaves, or open car parts, which are essentially thin shells with boundaries.

To represent these open surfaces, researchers have turned to Unsigned Distance Fields (UDFs). A UDF simply stores the distance to the nearest surface point, regardless of which side you are on. However, UDFs introduce a new challenge: because the distance is always non-negative, you cannot use standard algorithms like Marching Cubes, which look for a sign change to find the surface.

The Gradient Flip Intuition

The key insight that makes UDFs usable for meshing is not the distance value itself, but its gradient.

Imagine standing near a sheet of paper. As you move toward it, the distance decreases. Once you pass through the paper and move away from it on the other side, the distance begins to increase again. Crucially, the direction of the "closest point" has reversed. In a UDF, the gradient vector always points directly away from the nearest surface. Therefore, when you cross a surface, the gradient direction flips by 180 degrees.

Benoît Guillard, Federico Stella, and Pascal Fua leveraged this in MeshUDF by introducing pseudo-signs. By calculating the dot product of gradients at adjacent grid corners, they can detect surface intersections: if the dot product is negative, the gradients point in opposite directions, indicating a surface lies between them. This allows traditional sign-based meshing algorithms to be applied to non-watertight geometries.

Overcoming Non-Differentiability

While the gradient flip is a powerful indicator, UDFs are mathematically non-differentiable exactly at the surface (the zero level set), where the gradient direction is undefined or discontinuous. This often leads to "shattered" or fragmented mesh reconstructions.

To solve this, Baorui Ma, Zhizhong Han, and colleagues proposed Level Set Projection in Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection. Instead of relying on the noisy gradients at the surface, they use the stable, smooth gradients from points near the surface to "pull" query points onto the zero level set. This ensures surface continuity and high-fidelity shapes even when the field is difficult to optimize directly at the boundary.

Direct Intersection Prediction

Another evolution in this field moves away from interpreting the distance field entirely. In Surface Reconstruction from Point Clouds via Grid-based Intersection Prediction, Hui Tian and Kai Xu suggest that rather than regressing a UDF and then extracting a mesh, a model can directly predict the intersection points between grid edges and the surface. This bypasses the gradient discontinuity issues of UDFs altogether, producing smoother meshes for both open and closed surfaces.

Similarly, Junhui Hou and Ying He developed GeoUDF, which uses an "Edge-based Marching Cubes" algorithm. By analyzing the relationship between distance values and gradient directions specifically along grid edges, they can accurately localize boundaries in complex internal structures and open surfaces where traditional volumetric methods fail.

Applications in Rendering and Complex Topology

The shift to UDFs has enabled new capabilities in neural rendering. Xiaoxiao Long and Lingjie Liu introduced NeuralUDF, a framework that reconstructs surfaces with arbitrary topologies directly from 2D images. By using a gradient-aware volume rendering scheme, they can identify surface intersections during the ray-casting process by monitoring where the distance gradient abruptly reverses.

For even more complex cases, such as T-junctions or intersecting planes (non-manifold geometry), Xuhui Chen and Ying He developed MIND. This approach reframes reconstruction as a spatial partitioning problem, treating the surface as an interface between multiple "material" regions. By using local indicator functions to distinguish between sides of a surface patch, they avoid the "double-layer" artifacts that often plague simpler UDF-to-mesh conversions.

By moving from global signs to local gradient analysis, this research direction has successfully extended the flexibility of implicit representations to the messy, non-watertight reality of the physical world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_meshing_non_watertight_surfaces_from_uns_64954d61,
  title = {Meshing Non-Watertight Surfaces from Unsigned Distance Fields},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/meshing-non-watertight-surfaces-from-unsigned-distance-fields/rdr_64954d614d8f4904a8e158835f1d5daa},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_meshing_non_watertight_surfaces_from_uns_64954d61_v1,
  title = {Meshing Non-Watertight Surfaces from Unsigned Distance Fields},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/meshing-non-watertight-surfaces-from-unsigned-distance-fields/rdr_64954d614d8f4904a8e158835f1d5daa/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}