Research map / 3D vision / Face and shape
High Resolution UDF Meshing via Iterative Networks
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 an iterative neural network that meshes neural unsigned distance fields by repeatedly refining per-cell pseudo-sign predictions using previously estimated neighboring cell signs, improving high-resolution UDF meshing over single-pass methods.
Problem
Meshing neural unsigned distance fields (UDFs) into explicit surfaces is hard because UDFs have no sign changes and become noisy at high resolution, causing holes and missing surface in single-pass extraction methods.
Approach
A per-cell fully connected network predicts pseudo-sign configurations and is run iteratively, feeding its sigmoid-activated outputs for neighboring cells back as input, with random number of iterations and Gaussian noise augmentation during training, so that neighborhood context propagates across the grid.
Key results
- At resolution 512 on ShapeNet cars with an auto-decoder and Marching Cubes, the method achieves median CD 8.84e-5 versus 56.9e-5 for NSD-UDF + MC.
- Ours + DualMesh-UDF achieves the best median CD at 512 on MGN* (0.899 vs 1.13 for NSD-UDF + DualMesh-UDF) and on ShapeNet cars (13.5 vs 86.3 in the DC table).
- On point-cloud-trained CAP-L cars at 512 with MC, the method reaches median CD 10.5 versus 12.3 for NSD-UDF + MC, while being comparable to baselines at 128 resolution.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Gaussian noise augmentation during training (ablation) | iterative meshing network on ShapeNet cars auto-decoder · ShapeNet cars at 128³, 256³, and 512³ | CD × 10^-5: 8.01 -> 5.64; 12.5 -> 5.23; 47.8 -> 8.84 | Helped |
| train with 5 additional iterations rather than 1 (ablation) | iterative meshing network · MGN, ShapeNet cars, chairs, and planes at resolution 512 | Image Consistency: MGN: 94.9 -> 94.8; Cars: 88.5 -> 88.9; Chairs: 86.2 -> 87.2; Planes: 87.0 -> 87.1 | Mixed |
| use sigmoid activation for iterative outputs instead of identity (ablation) | iterative meshing network on ShapeNet cars · ShapeNet cars at resolution 512 | Image Consistency: 87.0 -> 88.6; 87.6 -> 88.5; 88.5 -> 88.9 | Helped |
| replace sigmoid activation with softmax (ablation) | iterative meshing network on ShapeNet cars · ShapeNet cars at resolution 512 | Image Consistency: 87.3 -> 87.5 | No clear effect |
| remove the learned iterative refinement and use NSD-UDF + MC | ShapeNet cars auto-decoder meshing · ShapeNet cars, 512-MC | CD × 10^-5: 8.84 -> 56.9 | Hurt |
| filter cells using UDF <0.1 and low confidence rather than no filtering (ablation) | proposed pipeline · ShapeNet cars at resolution 256 | Inference time: 7m -> 30s | Helped |
| increase iterations from 1 to 6 on ground-truth UDFs (ablation) | proposed meshing method · ShapeNet cars at resolution 512 | CD × 10^-5: 0.0294 -> 0.0300 | Hurt |
Surprising findings
- Removing the thresholds of MeshUDF and NSD-UDF improves high-resolution results but is insufficient alone, and MeshUDF can become less accurate at 256 when thresholds are removed.
- On ground-truth UDFs, multiple iterations provide no benefit, and the method introduces micro-holes comparable to other MC-based baselines.
Limitations
- Performance gains are small or absent at low resolutions and on simple shapes such as MGN, where existing methods already perform well.
- Mesh outputs still show blockiness and micro-holes with MC, and the method produces many small holes on ground-truth UDFs, without faithfully recovering topology.
Experimental setup
Datasets: ABC ShapeNet MGN 3D scene dataset (CAP-UDF repository) CAP-L point clouds DiffUDF cars
Benchmarks: ShapeNet cars ShapeNet chairs ShapeNet planes MGN CAP-L scenes CAP-L cars DiffUDF cars
Metrics: L2 Mesh Chamfer Distance F1 score Image Consistency
Baselines: CAP-UDF MeshUDF DCUDF DCUDF-T NSD-UDF + MC UNDC DualMesh-UDF DualMesh-UDF-T NSD-UDF + DualMesh-UDF
Builds on: NSD-UDF DualMesh-UDF CAP-UDF DeepSDF
Closest papers by meaning
- MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction (NeurIPS 2025 · similarity 0.91)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.90)
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.89)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.89)
- Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes (CVPR 2024 · similarity 0.88)
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.87)
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.