GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting

Shujuan Li, Yu-Shen Liu, Zhizhong Han

Abstract

Reconstructing open surfaces from multi-view images is vital in digitalizing complex objects in daily life. A widely used strategy is to learn unsigned distance functions (UDFs) by checking if their appearance conforms to the image observations through neural rendering. However, it is still hard to learn continuous and implicit UDF representations through 3D Gaussians splatting (3DGS) due to the discrete and explicit scene representation, i.e., 3D Gaussians. To resolve this issue, we propose a novel approach to bridge the gap between 3D Gaussians and UDFs. Our key idea is to overfit thin and flat 2D Gaussian planes on surfaces, and then, leverage the self-supervision and gradient-based inference to supervise unsigned distances in both near and far area to surfaces. To this end, we introduce novel constraints and strategies to constrain the learning of 2D Gaussians to pursue more stable optimization and more reliable self-supervision, addressing the challenges brought by complicated gradient field on or near the zero level set of UDFs. We report numerical and visual comparisons with the state-of-the-art on widely used benchmarks and real data to show our advantages in terms of accuracy, efficiency, completeness, and sharpness of reconstructed open surfaces with boundaries.

Introduction

High-fidelity 3D reconstruction from multi-view images typically relies on Signed Distance Functions (SDFs). While SDFs are highly effective for "watertight" objects—where a clear distinction exists between the interior and exterior—they struggle with "open" surfaces such as clothing, leaves, or thin sheets of paper. For these geometries, Unsigned Distance Functions (UDFs) are the preferred representation, as they simply measure the distance to the nearest surface without requiring a defined volume.

Historically, UDF methods have been implemented via Neural Radiance Fields (NeRFs). However, NeRF-based UDFs suffer from significant computational overhead due to expensive volume rendering and often produce noisy results because UDF gradients become mathematically unstable near the surface (the zero-level set). Recently, 3D Gaussian Splatting (3DGS) has emerged as a high-speed alternative for explicit scene representation. GaussianUDF bridges the gap between the discrete, explicit nature of 3DGS and the continuous, implicit nature of UDFs to create a framework that is both fast and accurate.

Methodology

GaussianUDF proposes a hybrid optimization framework that jointly learns a set of 2D Gaussians and a neural UDF network. This approach leverages the speed of Gaussian rasterization while maintaining the topological flexibility of implicit fields.

1. 2D Gaussian "Surfels"

Rather than using standard 3D ellipsoidal Gaussians, the framework utilizes thin, flat 2D Gaussians. These act as local surface elements, or "surfels," which are more naturally suited to represent thin structures and open boundaries. These Gaussians are overfitted to the surface to provide a geometric anchor for the implicit field.

2. Dual-Level Supervision

A core innovation of GaussianUDF is its approach to supervising the UDF field at different distances from the surface:

  • Far-Field Supervision ($L_{far}$): For points distant from the surface, the model uses the UDF's gradient field to guide query points toward the Gaussian centers, establishing the coarse global geometry.
  • Near-Field Supervision ($L_{near}$): Near the zero-level set, UDF gradients are notoriously unreliable. GaussianUDF bypasses this by sampling points along the normal vectors of the 2D Gaussians. Because the Gaussians are already aligned with the surface, their normals provide a stable geometric reference to supervise the UDF locally, ensuring a clean and sharp reconstruction.

3. Projection and Regularization

To maintain consistency between the discrete Gaussians and the implicit UDF, the framework employs a projection constraint ($L_{proj}$) that maps Gaussian centers onto the UDF's zero-level set. Additionally, depth distortion and normal alignment losses are used to ensure the Gaussians form a cohesive, smooth surface rather than a fragmented cloud.

Experimental Evaluation

The authors evaluated GaussianUDF across several diverse datasets, including DeepFashion3D (complex garments), the DTU dataset (watertight objects), and various real-world scans of plants and household items.

Quantitative Results

On the DeepFashion3D benchmark, GaussianUDF achieved a mean Chamfer Distance (CD) of 1.60, surpassing the previous state-of-the-art method, VRPrior (1.71). This indicates a higher degree of geometric fidelity in capturing complex folds and open boundaries.

Efficiency

One of the most significant advantages of the GaussianUDF framework is its efficiency. By utilizing Gaussian rasterization instead of NeRF-style ray tracing, the training time is reduced from over 9 hours (for methods like VRPrior) to approximately 1.6 hours. This represents a massive leap in practical usability for high-quality reconstruction.

Qualitative Comparison

Visual inspections demonstrate that GaussianUDF captures significantly sharper boundaries and finer details compared to existing methods like 2S-UDF or NeuralUDF. While previous methods often result in over-smoothed surfaces or artifacts near open edges, GaussianUDF maintains the integrity of thin structures and complex topology.

Conclusion and Future Work

GaussianUDF demonstrates that 3D Gaussian Splatting and Unsigned Distance Functions are complementary rather than competing technologies. By using 2D Gaussians as geometric anchors, the method avoids the gradient instability issues that have long plagued UDF optimization.

While the method excels in speed and detail, it faces challenges in textureless regions where initial Gaussian placement is difficult. Furthermore, extracting a high-quality mesh from a UDF remains more complex than the standard Marching Cubes algorithm used for SDFs. Future research may focus on integrating more advanced normal estimation or specialized meshing algorithms to further refine the reconstruction of high-frequency details in complex, real-world environments.

Summary updated

Abstract

Reconstructing open surfaces from multi-view images is vital in digitalizing complex objects in daily life. A widely used strategy is to learn unsigned distance functions (UDFs) by checking if their appearance conforms to the image observations through neural rendering. However, it is still hard to learn continuous and implicit UDF representations through 3D Gaussians splatting (3DGS) due to the discrete and explicit scene representation, i.e., 3D Gaussians. To resolve this issue, we propose a novel approach to bridge the gap between 3D Gaussians and UDFs. Our key idea is to overfit thin and flat 2D Gaussian planes on surfaces, and then, leverage the self-supervision and gradient-based inference to supervise unsigned distances in both near and far area to surfaces. To this end, we introduce novel constraints and strategies to constrain the learning of 2D Gaussians to pursue more stable optimization and more reliable self-supervision, addressing the challenges brought by complicated gradient field on or near the zero level set of UDFs. We report numerical and visual comparisons with the state-of-the-art on widely used benchmarks and real data to show our advantages in terms of accuracy, efficiency, completeness, and sharpness of reconstructed open surfaces with boundaries.
BibTeX
@inproceedings{li2025gaussianudf,
  author = {Shujuan Li and Yu-Shen Liu and Zhizhong Han},
  title = {{GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  url = {https://openreview.net/forum?id=HSuoUlqC95},
}
Figure