Research map / Point clouds / Shape and surface completion
High-Fidelity Lightweight Mesh Reconstruction from Point Clouds
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 a curvature-adaptive meshing method (point-based surface queries, learned vertex generator, Delaunay tetrahedral meshing) plus a grid+triplane hybrid SDF network to produce high-fidelity lightweight meshes from point clouds.
Problem
Marching Cubes on learned SDFs yields redundant mesh elements, while existing neural meshing methods on low-resolution grids lose sharp detail, have uniform resolution, and often fail to guarantee correct topology.
Approach
Project spatial queries onto the implicit surface via SDF gradients, generate any specified number of curvature-adaptive vertices with a learned displacement network, then connect them with a Delaunay tetrahedral classification (multi-label voting plus neighborhood label constraint) to produce watertight, manifold meshes; SDF accuracy is improved by a hybrid voxel-grid and tri-plane feature repres
Key results
- On Thingi10K with 1/4 the elements of comparison methods, Ours-lite reaches CE 0.124 vs PoNQ 0.201 at similar element count, and the full method improves CE by 78% over PoNQ at low resolution.
- On the Stanford dataset, Ours with 5% of dense-mesh vertices (AM5%) reaches CD 0.560e-5 and F1 0.959, outperforming dense-mesh baselines such as NeuralPull (CD 0.606e-5, F1 0.950).
- On SRB, the method achieves CD 2.561e-5 and NC 0.964 versus 4.004e-5 and 0.942 for the best explicit triangulation baseline (PointTriNet with FPS), with correct topology 1.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use curvature-adaptive meshing instead of the compared low-resolution meshing methods at the same element count | Lightweight mesh reconstruction · Thingi10K, grid resolution 32 | CE (10−3): 0.575 -> 0.126 | Helped |
| Use adaptive meshing with 5% of the dense mesh vertex count instead of NeuralPull dense reconstruction | SDF-based reconstruction · Stanford | CE (10−3): 0.329 -> 0.154 | Helped |
| Use adaptive meshing rather than point-cloud downsampling and triangulation | Lightweight mesh reconstruction · SRB | CD (10−5): 4.004 -> 2.561 | Helped |
| Add normal-consistency loss Lnc (ablation) | Vertex generator · Thingi10K | CD (10−5): 0.9643 -> 0.8551 | Helped |
| Add curvature-weighted loss Lcur (ablation) | Vertex generator · Thingi10K | CD (10−5): 0.9643 -> 0.8147 | Helped |
| Add both Lnc and Lcur (ablation) | Vertex generator · Thingi10K | CD (10−5): 0.9643 -> 0.7989 | Helped |
| Add progressive upsampling (ablation) | Vertex generator · Thingi10K | CD (10−5): 0.7550 -> 0.7550 | Helped |
| Use grid features only instead of no explicit features (ablation) | SDF network · Stanford | CD (10−5): 0.5894 -> 0.8282 | Hurt |
| Use grid features only instead of no explicit features (ablation) | SDF network · Stanford | CE (10−3): 0.3052 -> 0.1269 | Helped |
| Use tri-plane features only instead of no explicit features (ablation) | SDF network · Stanford | CD (10−5): 0.5894 -> 0.5639 | Helped |
| Combine grid and tri-plane features instead of using neither (ablation) | SDF network · Stanford | CE (10−3): 0.3052 -> 0.1059 | Helped |
| Combine grid and tri-plane features instead of using neither (ablation) | SDF network · Stanford | CD (10−5): 0.5894 -> 0.5580 | Helped |
Limitations
- Vertex generator requires per-object optimization (displacement network iterates 6k times per object)
- Relies on GT SDF for comparison baselines; main method learns SDFs from point clouds only
- Multi-label voting uses 101 reference points per tetrahedron, adding meshing cost
Experimental setup
Datasets: Stanford Thingi10K SRB ScanNet
Benchmarks: Thingi10K Stanford SRB ScanNet
Metrics: chamfer distance normal consistency F1-score curvature error correct topology
Baselines: Marching Cubes DMTet NDC NMC VoroMesh PoNQ NeuralPull PINC DiGS GridPull
Builds on: Neural Pulling Marching Cubes Point Transformer V3
Closest papers by meaning
- MultiPull: Detailing Signed Distance Functions by Pulling Multi-Level Queries at Multi-Step (NeurIPS 2024 · similarity 0.87)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.86)
- A Lightweight UDF Learning Framework for 3D Reconstruction Based on Local Shape Functions (CVPR 2025 · similarity 0.86)
- Adaptive 3D Reconstruction via Diffusion Priors and Forward Curvature-Matching Likelihood Updates (NeurIPS 2025 · similarity 0.86)
- SPU-PMD: Self-Supervised Point Cloud Upsampling via Progressive Mesh Deformation (CVPR 2024 · similarity 0.86)
- MIND: Material Interface Generation from UDFs for Non-Manifold Surface Reconstruction (NeurIPS 2025 · similarity 0.86)
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.