Research map / 3D vision / Face and shape
Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DEUDF, an unsupervised UDF learning method for 3D surface reconstruction from unoriented point clouds that combines an unconditioned SIREN MLP, normal alignment, adaptively weighted Eikonal loss, and DCUDF surface extraction.
Problem
Learning accurate unsigned distance fields (UDFs) for high-fidelity 3D surface reconstruction from unoriented point clouds, where UDFs suffer from non-differentiability at zero, vanishing gradients, many local minima, and oscillating gradient directions near the surface.
Approach
Relaxes the non-negativity constraint by using an unconditioned SIREN MLP to output signed-ish distances with a positivity term, aligns UDF gradients with PCA normals at points displaced on both sides of the surface, down-weights the Eikonal loss near the surface, and extracts the surface with DCUDF as the local minimal-distance surface.
Key results
- On Stanford 3D Scene, DEUDF achieves mean Chamfer-L1 3.09 x10^-3 and F1^0.005 99.41, better than CAP-UDF (3.37, 98.96), DUDF (3.79, 97.33) and LevelSetUDF (3.16, 99.17).
- On ShapeNet-Cars, DEUDF reaches mean Chamfer-L1 4.91 x10^-3 and F1^0.005 95.53, outperforming the UDF baselines.
- On noisy point clouds (N(0,0.0025)), DEUDF gets Chamfer mean 4.07 vs DUDF 4.27, but under N(0,0.005) DUDF is better (4.61 vs 5.64).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply absolute value to the SIREN output instead of using an unconditioned output (ablation) | DEUDF · Stonewall, Stanford 3D Scene | CD-mean: 2.69 -> 5.27 | Hurt |
| Apply softplus to the SIREN output instead of using an unconditioned output (ablation) | DEUDF · Stonewall, Stanford 3D Scene | F1^0.0025: 93.52 -> 91.08 | Hurt |
| Remove adaptive weighting from the Eikonal constraint (ablation) | DEUDF · Stonewall, Stanford 3D Scene | CD-mean: 2.69 -> 2.87 | Hurt |
| Remove normal alignment (ablation) | DEUDF · Stonewall, Stanford 3D Scene | CD-mean: 2.69 -> 3.20 | Hurt |
| Use random vectors instead of estimated normals for alignment (ablation) | DEUDF · Stonewall, Stanford 3D Scene | CD-mean: 2.69 -> 2.72 | Hurt |
Surprising findings
- Using fully random normals in the normal alignment loss still improves reconstruction significantly, suggesting the loss also helps locate the local minimal distance surface between paired samples.
- Despite relaxing non-negativity, DEUDF's average deviation from zero (1.34e-4) is smaller than that of strictly non-negative UDF methods such as LevelSetUDF (1.39e-4) and CAP-UDF (2.63e-4).
Limitations
- Designed for dense, evenly sampled point clouds; sparse or non-uniform inputs can produce holes.
- Sensitive to large noise: DUDF outperforms DEUDF at higher Gaussian noise levels, and SIREN frequency must be lowered for noisy inputs.
- Relies on PCA-estimated normals and a per-shape optimization of about 30 minutes per UDF.
Experimental setup
Datasets: ShapeNet-Cars Stanford 3D Scene Stanford 3D Scan
Benchmarks: Stanford 3D Scene Stanford 3D Scan ShapeNet-Cars
Metrics: Chamfer-L1 F-score average deviation from zero
Baselines: CAP-UDF DUDF LevelSetUDF NSH IDF NDF GIFS DeepSDF
Builds on: DCUDF SIREN IDF NeuralUDF LevelSetUDF
Closest papers by meaning
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.91)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.91)
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.89)
- MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction (NeurIPS 2025 · similarity 0.89)
- GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.