Research map / Point clouds / Shape and surface completion
MIND: Material Interface Generation from UDFs for Non-Manifold 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 MIND, which builds material interfaces directly from learned UDFs via a local two-signed side field, a global multi-labeled field, and multi-label Marching Cubes to extract non-manifold meshes.
Problem
Extracting correct non-manifold meshes (and open surfaces) from unsigned distance fields, where learned UDFs rarely attain exact zero distances and local SDF-based Marching Cubes produces holes, spurious components, or cannot represent non-manifold geometry.
Approach
Derive a material interface (multi-labeled spatial partition) from the UDF by computing a locally-windowed generalized-winding-number side field, cleaning it via erosion, connected-component labeling, alpha-expansion relabeling, and envelope-based merging, then extract the non-manifold mesh with multi-label Marching Cubes (M3C) and refine it against the UDF with a per-material-group Laplacian.
Key results
- On ShapeNet-Car (CAP-UDF UDFs), MIND achieves mean Chamfer 2.917 vs 3.086 for DMUDF and 3.290 for DCUDF.
- On DeepFashion3D (DCUDF UDFs), MIND achieves mean Chamfer 1.770 vs 1.792 for DMUDF and 1.825 for DCUDF.
- Qualitatively, MIND preserves non-manifold topology (checked by geodesic distances) where DCUDF yields double-layered meshes and NU-NeRF/NeUDF fail on transparent objects.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MIND instead of DMUDF for mesh extraction | CAP-UDF reconstructions · ShapeNet-Car | Chamfer distance L2 (Mean): 3.086 -> 2.917 | Helped |
| Use MIND instead of DCUDF for mesh extraction | CAP-UDF reconstructions · ShapeNet-Car | Chamfer distance L2 (Mean): 3.290 -> 2.917 | Helped |
| Use MIND instead of DMUDF for mesh extraction | DCUDF reconstructions · DeepFashion3D | Chamfer distance L2 (Mean): 1.792 -> 1.770 | Helped |
| Use MIND instead of DCUDF for mesh extraction | DCUDF reconstructions · DeepFashion3D | Chamfer distance L2 (Mean): 1.825 -> 1.770 | Helped |
| Use a smaller point-cloud downsampling voxel size (ablation) | MIND · UDF mesh reconstruction | reconstruction quality | No clear effect |
| Increase point-cloud downsampling voxel size to 0.01 (ablation) | MIND · UDF mesh reconstruction | reconstruction quality: 0.01 | Hurt |
| Reduce r1 to 0.03 (ablation) | MIND · UDF mesh reconstruction | reconstruction success: 0.03 | Hurt |
| Increase r2 to 0.02 (ablation) | MIND · UDF mesh reconstruction | region merging: 0.02 | Hurt |
| Reduce r2 to 0.0025 (ablation) | MIND · UDF mesh reconstruction | region merging: 0.0025 | Hurt |
Surprising findings
- DMUDF frequently produces non-manifold edges in regions that should be manifold, and DCUDF's double-layer meshes cause rendering artifacts.
- The method fails on non-orientable non-manifold surfaces where the two sides are indistinguishable.
Limitations
- Alpha-expansion labeling over the voxel grid is time-consuming.
- Erosion may remove small or thin material interface regions, causing missing facets.
- Cannot handle non-orientable surfaces where the two sides are indistinguishable; the method operates on pre-learned UDFs rather than learning MIs from raw data.
Experimental setup
Datasets: ShapeNet-Car DeepFashion3D
Benchmarks: ShapeNet-Car DeepFashion3D
Metrics: Chamfer distance L2 geodesic distance (topology check via heat method)
Baselines: DCUDF DMUDF NU-NeRF NeUDF MeshUDF
Builds on: M3C generalized winding number DCUDF alpha-expansion
Closest papers by meaning
- High Resolution UDF Meshing via Iterative Networks (NeurIPS 2025 · similarity 0.91)
- Details Enhancement in Unsigned Distance Field Learning for High-fidelity 3D Surface Reconstruction (AAAI 2025 · similarity 0.89)
- DUDF: Differentiable Unsigned Distance Fields with Hyperbolic Scaling (CVPR 2024 · similarity 0.89)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.88)
- Unsigned Orthogonal Distance Fields: An Accurate Neural Implicit Representation for Diverse 3D Shapes (CVPR 2024 · similarity 0.88)
- S2MDF: A Plug-And-Play Layer for Intersection-Free Multi-Object Signed Distance Fields (NeurIPS 2026 · 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.