Research map / Point clouds / Shape and surface completion
Learning Delaunay Surface Elements for Mesh Reconstruction
Marie-Julie Rakotosaona, Paul Guerrero, Noam Aigerman, Niloy J. Mitra, Maks Ovsjanikov
Check your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Delaunay surface elements with learned log maps, alignment, and triangle selection rather than the strongest baseline RVE | Mesh reconstruction · FamousThingi test set | NW (%): 9.2 -> 0.4 | Improved |
| Use Delaunay surface elements with learned log maps rather than PointTriNet | Mesh reconstruction · FamousThingi test set | CD *1^{e-2}: 0.337 -> 0.326 | Improved |
| Use the proposed method rather than IER meshing | Mesh reconstruction · ShapeNet, 100 random shapes | NW (%): 6.96 -> 5.51 | Improved |
| Add log-map alignment (ablation) | Ours · FamousThingi test set | NW (%): 1.07 -> 0.40 | Improved |
| Add triangle selection (ablation) | Ours · FamousThingi test set | NW (%): 10.98 -> 0.40 | Improved |
| Add both log-map alignment and triangle selection (ablation) | Ours · FamousThingi test set | NW (%): 22.51 -> 0.40 | Improved |
| Use learned log maps instead of tangent-plane projections (ablation) | Ours · FamousThingi test set | NW (%): 1.18 -> 0.40 | Improved |
| Use learned log maps instead of tangent-plane projection approximation (ablation) | Local parameterization · FamousThingi test set, 33 manifold shapes | Geodesic distance MSE *1^{e-3}: 1.943 -> 0.471 | Improved |
| Use learned log maps instead of the rotation-based approximation (ablation) | Local parameterization · FamousThingi test set, 33 manifold shapes | 2D position MSE *1^{e-2}: 2.835 -> 0.835 | Improved |
| Set geodesic neighbor count k to 30 rather than 20 (ablation) | Ours · FamousThingi test set | NW (%): 4.381 -> 0.485 | Improved |
| Set geodesic neighbor count k to 50 rather than 30 (ablation) | Ours · FamousThingi test set | NW (%): 0.485 -> 0.667 | Hurt |
| Set Euclidean neighbor count K to 120 rather than 80 at k=30 (ablation) | Ours · FamousThingi test set | NW (%): 0.509 -> 0.485 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PointTriNet — PointTriNet: Learned Triangulation of 3D Point Sets (ECCV 2020) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- High-Fidelity Lightweight Mesh Reconstruction from Point Clouds (CVPR 2025 · similarity 0.91)
- Meshing Point Clouds with Predicted Intrinsic-Extrinsic Ratio Guidance (ECCV 2020 · similarity 0.89)
- Surface Reconstruction From Point Clouds: A Survey and a Benchmark (TPAMI 2024 · similarity 0.89)
- Guaranteed Simply Connected Mesh Reconstruction from an Unorganized Point Cloud (ICLR 2026 · similarity 0.89)
- CircNet: Meshing 3D Point Clouds with Circumcenter Detection (ICLR 2023 · similarity 0.89)
- DMNet: Delaunay Meshing Network for 3D Shape Representation (ICCV 2023 · 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.
