Unsigned Distance Fields for Open Surface Reconstruction

233 concepts154 papers3 proposals

In 3D computer vision and graphics, implicit neural representations have fundamentally changed how we digitize the physical world. Instead of storing a shape as a discrete collection of voxels or polygons, networks learn a continuous function that maps any 3D coordinate to a geometric property.

For years, the dominant formulations were Occupancy Networks and Signed Distance Functions (SDFs). An SDF maps a 3D point to its closest distance to a surface, using a positive sign for points outside the object and a negative sign for points inside. The physical surface is implicitly defined as the "zero level set"—the boundary where the sign flips from positive to negative.

While mathematically elegant, this formulation harbors a fatal assumption: the object must have a clearly defined "inside" and "outside." In other words, the geometry must be a closed, watertight manifold.

As noted by Cheng Lin et al. in SparseNeuS: Fast Generalizable Neural Surface Reconstruction from Sparse Views, this volumetric requirement makes SDFs inherently incapable of modeling open surfaces. Real-world environments are full of geometries that violate this constraint: a single leaf, a piece of clothing with functional openings (necklines and armholes), an unzipped jacket, or the complex internal structures of a car interior. When forced to represent these shapes, SDFs either artificially close the boundaries, create distorted geometries, or cause the optimization process to collapse entirely.

Shifting to Unsigned Distance Fields

To break the watertight manifold constraint, researchers shifted from signed to unsigned representations. In Neural Unsigned Distance Fields for Implicit Function Learning, Julian Chibane et al. introduced Neural Unsigned Distance Fields (UDFs).

The core intuition behind a UDF is simple: drop the sign. Instead of trying to determine whether a point is inside or outside a shape, a UDF predicts only the absolute, unsigned distance $d \ge 0$ to the closest surface point.

Signed Distance Field (SDF)       Unsigned Distance Field (UDF)
      Outside (+)                       Anywhere (d > 0)
  -------------------                 -------------------
      Inside (-)                        Anywhere (d > 0)
  (Requires Closed Volume)            (Supports Open Sheets)

Because it does not rely on a spatial sign flip, a UDF can represent any arbitrary topology. An open sheet of paper or a complex garment can be represented because the surface is simply the locus of points where $d = 0$. This flexibility has enabled generative models like Surf-D: Generating High-Quality Surfaces of Arbitrary Topologies Using Diffusion Models by Zhengming Yu et al. and FullFormer: Generating Shapes Inside Shapes to synthesize complex, non-watertight objects and nested internal structures that were previously impossible to capture implicitly.

The Challenge of the Zero Level Set

If UDFs are so flexible, why haven't they completely replaced SDFs? Dropping the sign introduces two severe mathematical hurdles: non-differentiability and the lack of sign-crossings for mesh extraction.

1. Non-Differentiability at the Surface

Consider the absolute value function $f(x) = |x|$. At $x = 0$, the function forms a sharp V-shape where its derivative is undefined; the gradient abruptly flips direction.

A UDF suffers from this exact problem at the surface. As an optimization path approaches an open surface from one side, the distance decreases toward zero. The moment it crosses the surface, the distance immediately begins to increase again. Consequently, the gradient of a UDF is discontinuous precisely at the zero level set ($d=0$). Because neural networks rely on smooth gradients to learn, this discontinuity makes training highly unstable near the surface, frequently yielding "shattered," noisy, or fragmented geometric boundaries.

2. No Sign Crossings for Mesh Extraction

Standard mesh extraction algorithms, such as Marching Cubes, evaluate a grid of implicit values and look for edges where the values cross from negative to positive. Because a UDF is strictly non-negative ($d \ge 0$), traditional sign-crossing algorithms cannot locate the surface.

Smoothing the Gradient with Vector Fields

To resolve the gradient discontinuity at the surface, researchers realized they needed to predict more than just a scalar distance. Instead of asking "how far away is the surface?", networks can predict a vector pointing directly to the closest surface point.

Rahul Venkatesh et al. implemented this intuition in Deep Implicit Surface Point Prediction Networks through the Closest Surface-Point (CSP) representation. By training the network to output exact coordinates of the nearest surface point, the representation avoids the non-differentiable "flip" of scalar UDFs.

Similarly, Federico Stella et al. introduced the Gradient Distance Function (GDF). The GDF directly predicts the direction and distance to the surface as a continuous vector field. Because a vector field pointing toward a surface does not invert its components when passing through an open sheet, the function remains fully differentiable right at the surface boundary, eliminating the holes and fragmentation typical of raw UDF optimizations.

Scalar UDF Gradient Discontinuity       Vector Field (GDF / CSP) Smoothness
          |                                           |
     ---> | <---  (Gradients Clash)              ---> | --->  (Vectors Align)
          |                                           |

Another perspective is to avoid evaluating the function directly on the problematic zero level set during training. In Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection, Baorui Ma et al. proposed sampling points near the surface where the UDF gradients are still clean and stable. They developed a mathematical projection operator that uses these off-surface gradients to "pull" nearby queries onto the surface, ensuring a continuous and smooth zero level set without directly optimizing the discontinuous boundary.

Extracting Meshes Without Sign Changes

Even with a beautifully optimized UDF, converting that field into a standard polygonal mesh for rendering engines requires dedicated extraction strategies.

One prominent approach is the concept of "double covering." In Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering, Fei Hou et al. bypassed the lack of sign crossings by first extracting an iso-surface at a small distance $\epsilon > 0$ from the target. This creates a thin, watertight volumetric "shell" that encapsulates the open surface from both sides. They then use a graph-cut optimization to collapse this double-layered shell back down onto the true zero-distance surface, yielding a clean, single-layered open mesh.

  UDF Surface (d=0)          Dilated Shell (d=ε)         Collapsed Mesh
---------------------       =====================     ---------------------
                             (Watertight Double)        (Single-Layer Open)

Other approaches modify the underlying mesh generation framework. For example, Zhiqin Chen et al. introduced an unsigned variant of Dual Contouring in Neural Dual Contouring. Instead of relying on sign changes along grid edges, their method trains a network to predict vertex positions and edge crossings directly from sign-agnostic features, naturally preserving sharp geometric features on thin sheets.

When dealing with highly complex, non-manifold structures—such as T-junctions or intersecting internal walls—even double-covering can struggle. To solve this, Xuhui Chen et al. introduced MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction. Instead of treating the surface as a boundary between inside and outside, they reframe mesh extraction as a spatial partitioning problem. The algorithm treats the open or branching surface as an interface separating multiple distinct "material" regions in space, allowing the robust extraction of complex, intersecting open geometries without double-layer artifacts.

Alternatively, some frameworks attempt to unify the benefits of both worlds. In HSDF: Hybrid Sign and Distance Field for Modeling Surfaces with Arbitrary Topologies, Li Wang proposed predicting a distance field and a sign field independently, fusing them via gradient-based optimization to maintain the easy mesh extraction of SDFs where volumes exist, while gracefully falling back to UDF behavior for open structures.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 167 papers
  2. Version 1 · 2026-10-04 · 239 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_watertight_manifold_constraint_8165769e,
  title = {watertight manifold constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/watertight-manifold-constraint/rdr_8165769e0e794b558a06aea16e805d19},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_watertight_manifold_constraint_8165769e_v2,
  title = {watertight manifold constraint},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/watertight-manifold-constraint/rdr_8165769e0e794b558a06aea16e805d19/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake