Neural Surface Reconstruction for Open and Non-Watertight Geometries

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2239 papers277 conceptsfrozen 2026-10-04

In 3D computer vision, the dominant way to represent a shape implicitly has long been the Signed Distance Field (SDF). An SDF tells you how far any point in space is from a surface, with a crucial detail: the distance is negative if you are "inside" the object and positive if you are "outside." This simple sign bit allows algorithms to easily find the surface (the "zero level set") and enables efficient rendering and meshing.

However, this mathematical elegance comes with a strict requirement: the object must be watertight. It must have a clearly defined interior and exterior. As noted by Junkai Deng and Fei Hou in 2S-UDF, this constraint fails for many real-world objects. A t-shirt, a leaf, or a car interior are "open" surfaces—they are thin shells that do not enclose a volume. For these geometries, the concept of "inside" is meaningless, and traditional SDFs break down.

A growing body of research is shifting toward Unsigned Distance Fields (UDF) and alternative implicit representations to model these complex, non-manifold topologies.

The Shift from Signs to Proximity

The core intuition behind an Unsigned Distance Field is to discard the requirement for an interior. A UDF simply predicts the absolute distance to the nearest surface point. Because it doesn't need to partition space into "in" and "out," it can represent arbitrary geometries, including "triangle soups" and open boundaries.

In DUDE: Deep Unsigned Distance Embeddings, the authors demonstrate that by decoupling proximity (the UDF) from orientation (a Normal Vector Field), a model can represent complex open surfaces with much higher fidelity than a standard SDF. Similarly, Rahul Venkatesh and colleagues proposed the Closest Surface-Point (CSP) representation in Deep Implicit Surface Point Prediction Networks. Instead of just predicting a distance scalar, CSP predicts the actual coordinates of the nearest point on the surface, allowing for significantly more accurate normal estimation and faster rendering.

Overcoming the Gradient Discontinuity

If UDFs are so flexible, why haven't they replaced SDFs entirely? The challenge lies in optimization. In an SDF, the gradient is continuous as you cross the surface. In a UDF, the distance decreases as you approach the surface and increases as you move away, creating a "V-shape" profile. At the surface itself (distance = 0), the gradient is non-differentiable—it effectively "flips" direction instantaneously.

This makes it difficult for neural networks to learn a smooth, continuous surface. Without intervention, UDF-based reconstructions often appear "shattered" or fragmented. To solve this, Baorui Ma and Junsheng Zhou introduced Level Set Projection in Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection. Their method uses the reliable gradients of points near the surface to "pull" the zero level set into a cohesive shape, ensuring topological continuity.

Rendering and Meshing Open Surfaces

To make these representations useful, we must be able to visualize them (rendering) and convert them into usable formats (meshing).

  1. Neural Rendering: Traditional volume rendering assumes light travels through a volume. For thin, open surfaces, this leads to artifacts. Xiaoxiao Long and Cheng Lin developed NeuralUDF, which introduces a gradient-aware rendering scheme. By analyzing the UDF's gradient, the model can identify where a surface exists even without a sign change, enabling high-quality multi-view reconstruction of objects like clothing and leaves.
  2. Differentiable Meshing: Converting a UDF to a mesh is harder than the standard Marching Cubes algorithm used for SDFs. Benoît Guillard and Federico Stella addressed this with MeshUDF, an algorithm that uses the UDF's gradient field to extract surfaces. This allows for end-to-end training where the loss can be defined directly on the resulting mesh.

Applications in Generative Modeling

The ability to handle arbitrary topologies has opened new doors for 3D generative models. Traditional diffusion models for 3D often struggle with the "watertight" requirement of the datasets they are trained on. Surf-D, authored by Zhengming Yu and Zhiyang Dou, uses Latent Diffusion Models paired with UDFs to generate high-quality surfaces of arbitrary topologies. This allows for the synthesis of complex objects like open furniture or intricate garments that were previously impossible to represent accurately.

Similarly, GEOMDIST by Biao Zhang treats geometry as a continuous probability distribution of surface points. By moving away from fixed distance fields entirely, it can represent non-manifold geometries like hollow tubes or thin foliage with significantly fewer parameters.

Summary

The transition from Signed to Unsigned Distance Fields represents a fundamental shift in how we prioritize geometric flexibility. By relaxing the "watertight" constraint, researchers like Dmitriy Smirnov (DeepCurrents) and Huan Lei (CircNet) are enabling the digital reconstruction of the messy, open-ended geometry of the real world—from the folds of a garment in Deep Fashion3D to the complex boundaries of a large-scale indoor scene in RangeUDF.

Papers in this version

Showing 60 of 239 papers.

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 1, frozen text):

@misc{lacuna_neural_surface_reconstruction_for_open_a_8165769e_v1,
  title = {Neural Surface Reconstruction for Open and Non-Watertight Geometries},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-surface-reconstruction-for-open-and-non-watertight-geometries/rdr_8165769e0e794b558a06aea16e805d19/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}