Research map / Point clouds / Point cloud classification
Weighted Poisson-disk Resampling on Large-Scale 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 weighted Poisson-disk (WPD) resampling for large-scale point clouds, combining voxel-based Poisson radius estimation with weighted tangent smoothing (cotangent-weighted Voronoi-style displacement) and optional sharp-feature preservation.
Problem
Resampling large-scale point clouds to a specified point number with uniform density and geometric consistency; existing Poisson-disk radius estimation is inaccurate and FPS variants and feature-based methods are slow on large inputs.
Approach
Estimates surface area via voxel-based accumulation to set a more accurate Poisson-disk radius, iteratively refines the radius to hit the target count, then applies a cotangent-weighted tangent-space central displacement (CVT-inspired) to make distributions isotropic, with edge points given halved radii to keep sharp features.
Key results
- On urban scenes with 2,000K-200K input points and 10K outputs, WPD takes 12.7s versus >10 minutes for WLOP, Laplacian and AIVS.
- Poisson-disk resampling in MeshLab shows over 30% average error in controlling output point count, while WPD targets error under 5% via iterative refinement.
- Table of geometric consistency reports WPD achieving best Hausdorff in several scenes but not all; it is not uniformly best (e.g., Poisson wins scene646 and 653 Hausdorff).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use weighted Poisson-disk resampling instead of comparison methods | RGB-D point-cloud resampling · RGB-D scenes 645, 646, 653, 667, 673 | Voronoi density error: 33.96 -> 5.015; 27.42 -> 6.861; 25.41 -> 9.735; 21.11 -> 6.873; 36.73 -> 10.14 | Helped |
| Use weighted Poisson-disk resampling instead of Poisson resampling | RGB-D point-cloud resampling · RGB-D scene645 | Hausdorff distance: 421.62 -> 421.61 | No clear effect |
| Use weighted Poisson-disk resampling instead of Poisson resampling | RGB-D point-cloud resampling · RGB-D scenes 646 and 653 | Hausdorff distance: 313.19 -> 390.36; 317.60 -> 402.47 | Hurt |
| Use weighted Poisson-disk resampling instead of Poisson resampling | RGB-D point-cloud resampling · RGB-D scenes 667 and 673 | Hausdorff distance: 376.51 -> 327.42; 440.06 -> 432.34 | Helped |
| Use WPD for large-scale resampling instead of WLOP | Large-scale point-cloud resampling · 2,000K-200K input points; output 10k | resampling time: >$10min -> 12.7s | Helped |
| Use WPD for resampling | Large-scale point clouds · Resampling point-count control | resampling error: under $5\% | Helped |
| Use WPD instead of FPS, PU-Net, or default sampling for PointNet++ | PointNet++ semantic segmentation · 10 S3DIS scenes | mIoU | Helped |
Surprising findings
- Original Poisson-disk resampling in several comparisons attains lower Hausdorff distance than WPD on some scenes, so WPD is not uniformly best on geometric consistency.
Limitations
- Relies on an already fairly uniform input from the Poisson step; non-uniform inputs can cause slow convergence or negative cotangent weights.
- Sharp feature detection is limited by noise, and enhanced features are only partially recovered in noisy scenes.
- Semantic analysis evaluated only with PointNet++ on S3DIS over 10 scenes.
Experimental setup
Datasets: RGB-D dataset BuildingNet SensatUrban S3DIS UrbanScene3D UrbanBIS
Benchmarks: RGB-D dataset scenes BuildingNet S3DIS SensatUrban UrbanBIS
Metrics: Hausdorff distance mean distance local density error Voronoi density error runtime mIoU point count error
Baselines: Clustering WLOP Poisson-disk (MeshLab) Laplacian PU-Net AIVS FPS
Builds on: Poisson-disk resampling isotropic remeshing centroidal voronoi tessellation MSL-Net
Closest papers by meaning
- Arbitrary-Scale Point Cloud Upsampling by Voxel-Based Network with Latent Geometric-Consistent Learning (AAAI 2024 · similarity 0.87)
- SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network (AAAI 2025 · similarity 0.86)
- PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling (AAAI 2025 · similarity 0.86)
- A Conditional Denoising Diffusion Probabilistic Model for Point Cloud Upsampling (CVPR 2024 · similarity 0.85)
- GCD-Sampling: A General Cross-scale Decoupled Sampling for Point Cloud (AAAI 2025 · similarity 0.85)
- Point Cloud Upsampling Using Conditional Diffusion Module with Adaptive Noise Suppression (CVPR 2025 · similarity 0.84)
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.