Research map / Point clouds / Shape and surface completion
Parametric Point Cloud Completion for Polygonal 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.
Introduces parametric point cloud completion, where a transformer (PaCo) recovers plane primitives with inlier points from incomplete point clouds to enable polygonal surface reconstruction.
Problem
Polygonal surface reconstruction from incomplete point clouds fails because standard point cloud completion methods recover individual points rather than the parametric planar structures needed for polygonal surfaces.
Approach
Replace point-wise completion with parametric completion: hierarchically encode points into plane proxies (plane parameters plus inlier points), generate proxy proposals with a transformer, and train with bipartite matching over classification, normal, chamfer and repulsion losses so the model predicts variable-size planar primitives.
Key results
- PaCo achieves CD 1.87, HD 4.09, NC 0.943, FR 0.48 with PolyFit solver on ABC, beating AdaPoinTr (CD 3.16) and ODGNet (CD 2.73).
- On hard (75% missing) data PaCo reaches CD 2.43 vs. 4.72 for ODGNet and 5.81 for AdaPoinTr.
- Direct reconstruction baselines fail on many inputs (PolyFit FR 63.28%), and completion-plus-reconstruction outperforms direct neural reconstruction (SECAD-Net CD 4.35).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use parametric completion with plane proxies instead of conventional point completion | Polygonal surface reconstruction on ABC · ABC test, PolyFit | CD: 2.73 -> 1.87 | Helped |
| Use parametric completion with plane proxies instead of conventional point completion | Polygonal surface reconstruction on ABC · ABC test, COMPOD | CD: 3.22 -> 1.94 | Helped |
| Use parametric completion with plane proxies instead of conventional point completion | Polygonal surface reconstruction on ABC · ABC test, KSR | CD: 2.90 -> 1.91 | Helped |
| Use PaCo completion before polygonal reconstruction rather than direct reconstruction from incomplete input | Polygonal surface reconstruction on ABC · ABC test, direct reconstruction | CD: 4.35 -> 1.87 | Helped |
| Use PaCo rather than ODGNet on hard inputs | Point cloud completion and reconstruction · ABC test, incompleteness hard (-75%) | CD: 4.72 -> 2.43 | Helped |
| Replace polar coordinates with Cartesian coordinates (ablation) | PaCo · ABC test, architecture ablation | CD: 2.11 -> 3.34 | Hurt |
| Remove plane proxies (ablation) | PaCo · ABC test, architecture ablation | CD: 2.11 -> 2.45 | Hurt |
| Replace sum pooling with mean pooling (ablation) | PaCo · ABC test, architecture ablation | CD: 2.11 -> 2.31 | Hurt |
| Add repulsion loss (Lrep) (ablation) | PaCo · ABC test, loss ablation | CD: 2.48 -> 2.11 | Helped |
| Add overall chamfer loss (Lco) (ablation) | PaCo · ABC test, loss ablation | CD: 2.45 -> 2.11 | Helped |
| Add plane normal loss (Lnorm) (ablation) | PaCo · ABC test, loss ablation | CD: 2.22 -> 2.11 | Helped |
Surprising findings
- Direct polygonal reconstruction from incomplete data performs worse than the completion-then-reconstruction strategy, even for neural reconstruction methods.
- Removing the plane proxy stage or replacing sum pooling with mean pooling noticeably hurts performance.
Limitations
- Restricted to point clouds containing only planar primitives; may fail on non-planar geometry.
- Hierarchical encoding relies on planar structures being present in the input.
- Not optimized for conventional point cloud completion that prioritizes individual point distribution.
Experimental setup
Datasets: ABC
Benchmarks: ABC
Metrics: chamfer distance Hausdorff distance normal consistency failure rate
Baselines: PCN FoldingNet GRNet PoinTr AdaPoinTr ODGNet PolyFit KSR COMPOD BSP-Net SECAD-Net QEM
Builds on: PoinTr AdaPoinTr GoCoPP PolyFit Hungarian matching
Closest papers by meaning
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.86)
- SymmCompletion: High-Fidelity and High-Consistency Point Cloud Completion with Symmetry Guidance (AAAI 2025 · similarity 0.86)
- Learning Bijective Surface Parameterization for Inferring Signed Distance Functions from Sparse Point Clouds with Grid Deformation (CVPR 2025 · similarity 0.86)
- Point-based Instance Completion with Scene Constraints (ICLR 2025 · similarity 0.85)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.85)
- DC-PCN: Point Cloud Completion Network with Dual-Codebook Guided Quantization (AAAI 2025 · similarity 0.85)
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.