Research map / Point clouds / Registration and matching
GRICP: Granular-Ball Iterative Closest Point with Multikernel Correntropy for Point Cloud Fine Registration
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 GRICP, a robust, globally-aware ICP that converts point clouds into granular-ball clouds and uses multikernel correntropy as the loss to derive closed-form rigid transforms for point cloud fine registration.
Problem
ICP for point cloud fine registration is sensitive to outliers and tends to converge to local optima, especially under large rotation offsets, partial overlap, and Gaussian noise.
Approach
Replace point-level correspondences with granular-ball (coarse-grained ball) correspondences scored by center distance plus overlap ratio, filter outliers with ball radius and overlap-feature differences, and optimize the rigid transform by maximizing a multikernel correntropy objective solved via half-quadratic optimization and SVD.
Key results
- Under 100°-180° rotational offsets, GRICP reports rotation errors around 0.001–0.036 and translation errors 0.002–0.019, versus Go ICP's 0.005–0.082 rotation error and FGR's up to 1.05.
- On partial overlap (60%–90%) on Stanford, GRICP attains RE between 0.030 and 0.412 while standard ICP variants stay around 2.7–3.1 RE.
- GRICP runs in 0.10 s on Stanford and 2.86 s on 3DMatch, versus 4.21 s and 599.08 s for Go ICP on the two datasets respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRICP instead of GoICP | Point-cloud registration · Stanford 3D Scanning Repository and 3DMatch, large rotation offsets | RE: 0.031 -> 0.008; 0.082 -> 0.006; 0.005 -> 0.001 | Helped |
| Use GRICP instead of GoICP | Point-cloud registration · Stanford 3D Scanning Repository and 3DMatch, 150° rotation | RE: 0.012 -> 0.036 | Hurt |
| Use GRICP instead of GoICP | Point-cloud registration · Stanford 3D Scanning Repository and 3DMatch, large rotation offsets | TE: 0.011 -> 0.002; 0.013 -> 0.017; 0.014 -> 0.019 | Mixed |
| Use GRICP instead of strongest reported baseline | Point-cloud registration · Stanford 3D Scanning Repository, partial overlap | RE: 1.744 -> 0.412; 2.962 -> 0.308; 0.031 -> 0.143; 0.030 -> 0.030 | Helped |
| Use GRICP instead of strongest reported baseline | Point-cloud registration · Stanford 3D Scanning Repository, partial overlap | TE: 0.183 -> 0.125; 0.119 -> 0.145; 0.022 -> 0.015; 0.016 -> 0.015 | Mixed |
| Use GRICP instead of strongest reported baseline | Point-cloud registration · 3DMatch, Gaussian noise | RE: 0.349 -> 0.287; 0.208 -> 0.101; 0.044 -> 0.062; 0.176 -> 0.238 | Mixed |
| Use GRICP instead of strongest reported baseline | Point-cloud registration · 3DMatch, Gaussian noise | TE: 0.158 -> 0.150; 0.052 -> 0.046; 0.146 -> 0.127 | Helped |
| Use granular-ball correspondences instead of FPFH correspondences | Correspondence generation · 3DMatch, 10%–50% noise | precision: 0.2897 -> 0.3536; 0.2124 -> 0.3218; 0.1833 -> 0.2825; 0.1378 -> 0.2385 | Helped |
| Use granular-ball correspondences instead of FPFH correspondences | Correspondence generation · 3DMatch, no noise | precision: 0.4196 -> 0.3848 | Hurt |
| Use granular-ball correspondences instead of FPFH correspondences | Correspondence generation · 3DScanRepo, 30% and 50% noise | precision: 0.6526 -> 0.6716; 0.6027 -> 0.6346 | Helped |
Surprising findings
- Under 10%-50% Gaussian noise on 3DMatch, GRICP does not dominate: at 10% noise, ICP and RICP (RE 0.349) and Go ICP (RE 0.185) beat GRICP (RE 0.287), and at 50% noise GRICP's RE (0.238) is worse than AA ICP and Fast ICP.
Limitations
- Performance under heavy Gaussian noise is not consistently best; GRICP is outperformed by some baselines at low noise levels and by Go ICP on several settings.
- Correspondence generation via granular balls is less precise than FPFH on noise-free 3DMatch and 3DScanRepo data (Table 2), so it relies on ball-level matching.
- Experiments are limited to a few Stanford models and 3DMatch scenes, and evaluation hardware is a laptop CPU.
Experimental setup
Datasets: Stanford 3D Scanning Repository 3DMatch
Benchmarks: Stanford 3D Scanning Repository 3DMatch
Metrics: rotation error (RE) translation error (TE) run time (s) correspondence precision
Baselines: ICP Sparse ICP AA ICP Fast ICP Robust ICP Go ICP FGR
Builds on: Iterative Closest Point Maximum Correntropy Criterion MultiKernel Correntropy granular-ball computing Go-ICP
Closest papers by meaning
- Scalable and Differentiable Point-Cloud Registration Using Maximum Mean Discrepancy (ICML 2026 · similarity 0.87)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.86)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.85)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.85)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · 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.