Abstract
Reconstructing open surfaces from multi-view images is a significant challenge in 3D computer vision. While Signed Distance Functions (SDFs) excel at watertight objects, they struggle with open structures like clothing or leaves because these geometries lack a clear interior volume. Unsigned Distance Functions (UDFs) offer a solution but typically rely on slow NeRF-based volume rendering. We present GaussianUDF, a framework that bridges the gap between the speed of 3D Gaussian Splatting (3DGS) and the topological flexibility of UDFs. By overfitting 2D Gaussians onto surfaces and implementing a dual-zone supervision strategy, GaussianUDF achieves high-fidelity, continuous surface reconstruction with significantly reduced training times.
Introduction
Digitalizing complex, thin objects with boundaries—such as garments, foliage, and paper—requires representations that can handle "open" surfaces. Traditional implicit models like SDFs assume a clear interior and exterior, making them unsuitable for these geometries. UDFs, which represent the distance to the nearest surface without a sign, are better suited but have historically been slow to optimize due to their reliance on expensive ray-tracing and intersection finding.
Conversely, 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction with its speed and explicit representation. However, 3DGS produces discrete elements that do not naturally form a continuous, smooth surface. GaussianUDF resolves this by using 2D Gaussians as a bridge to supervise a continuous UDF field, combining the efficiency of splatting with the smoothness of implicit surfaces.
Methodology
The core of GaussianUDF lies in its ability to translate discrete 2D Gaussians into a continuous distance field through three primary mechanisms:
1. 2D Gaussian Splatting (2DGS)
Instead of 3D ellipsoids, the method utilizes 2D Gaussian primitives—essentially tiny, flat disks. These are inherently better at representing thin surfaces and provide a clear geometric basis for local distance calculations.
2. Dual-Zone Supervision
To train the UDF network, the system supervises distances in two distinct regions:
- Far Field: Query points far from the surface are "pulled" toward the Gaussian centers using a gradient-based inference method.
- Near Field: For points near the surface, the system leverages the flatness of 2D Gaussians. By sampling points along the normal vector of the Gaussian disks, the exact distance to the plane is known, providing a strong self-supervision signal to the UDF.
3. Stability and Consistency Constraints
To ensure the Gaussians accurately represent the underlying surface, a Projection Loss ($L_{proj}$) is introduced. This loss explicitly moves Gaussian centers onto the zero level set of the UDF, preventing optimization drift. Additionally, depth and normal consistency losses ensure the reconstruction aligns perfectly with the visual input from multi-view images.
Experimental Results
GaussianUDF was evaluated on benchmarks including DeepFashion3D (DF3D), the DTU dataset, and real-captured data.
Quantitative Performance
On the DF3D dataset, GaussianUDF achieved a Chamfer Distance of 1.60, significantly outperforming previous state-of-the-art UDF methods like 2S-UDF (1.98) and VRPrior (1.71). Crucially, training time was reduced from approximately 8-9 hours for NeRF-based competitors to just 1.6 hours.
Qualitative Comparison
Visual results demonstrate that GaussianUDF captures sharp boundaries and complex folds that other methods either blur or fail to reconstruct entirely. As shown in the comparison below, GaussianUDF maintains the integrity of thin structures like pinwheel petals and umbrella surfaces where baselines produce artifacts or fail to reconstruct the open geometry.
Conclusion
GaussianUDF successfully combines the efficiency of Gaussian Splatting with the flexibility of Unsigned Distance Functions. By bridging discrete and implicit representations, it enables the rapid reconstruction of complex open surfaces with high accuracy, completeness, and sharpness. Future work will focus on improving surface extraction (meshing) and handling textureless regions where Gaussians lack visual anchors.
Summary updated