Research map / Point clouds / Registration and matching
Progressive Correspondence Regenerator for Robust 3D 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 Regor, a progressive correspondence regenerator that iteratively generates new high-quality 3D point correspondences via prior-guided local grouping, generalized mutual matching, center-aware three-point consistency, and global refinement, rather than only pruning outliers.
Problem
Robust point cloud registration when initial correspondences contain extreme outlier ratios and very few inliers, where outlier removal methods leave too few correct correspondences to estimate the pose.
Approach
Instead of subtracting outliers from the initial set, the method progressively regenerates correspondences: each iteration samples seed correspondences, builds local regions around them with shrinking radius, re-matches features with generalized mutual matching, corrects local correspondences using a center-aware three-point consistency, and then merges and refines globally with second-order consi
Key results
- On 3DMatch with FPFH, Regor reaches 88.48% registration recall, 5.2 pp above SC2-PCR, and an INR of 1253.74% versus 78.78% for SC2-PCR.
- On the 3DMatch-EOR benchmark with outlier ratios above 99%, Regor achieves 26.76% RR while SC2-PCR achieves 9.15%.
- With RoPS features, Regor raises registration recall from 54.16% to 91.37%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| progressive regeneration instead of one-stage processing (ablation) | Regor · 3DMatch | RR: 86.81 -> 88.48 | Helped |
| progressive regeneration instead of one-stage processing (ablation) | Regor · 3DMatch | IN: 290.32 -> 2532.40 | Helped |
| progressive regeneration instead of one-stage processing (ablation) | Regor · 3DMatch-EOR | RR: 23.94 -> 26.76 | Helped |
| replace mutual matching with generalized mutual matching (ablation) | Regor · 3DMatch | RR: 86.81 -> 88.48 | Helped |
| replace nearest-neighbor matching with generalized mutual matching (ablation) | Regor · 3DMatch | RR: 87.35 -> 88.48 | Helped |
| add global correspondence refinement (ablation) | Regor · 3DMatch | RR: 87.35 -> 88.48 | Helped |
| add local correspondence refinement (ablation) | Regor · 3DMatch | RR: 76.37 -> 88.48 | Helped |
| use center-aware three-point consistency instead of second-order consistency (ablation) | Regor · 3DMatch | RR: 86.88 -> 88.48 | Helped |
| use center-aware three-point consistency instead of second-order consistency (ablation) | Regor · 3DMatch-EOR | RR: 22.54 -> 26.76 | Helped |
| add Regor to FPFH features | FPFH · 3DMatch | RR(%): 54.34 -> 88.48 (+34.14) | Helped |
| add Regor to RoPS features | RoPS · 3DMatch | RR(%): 54.16 -> 91.37 (+37.21) | Helped |
Surprising findings
- Outlier removal methods are capped at below 100% INR by construction, while the regenerator produces roughly 10 times more correct correspondences than the initial set.
Limitations
- Relies on seed correspondences from the previous stage being partially correct; errors in seeds can propagate into local refinement.
- Absolute RR on 3DMatch-EOR at >99% outlier ratio remains low (26.76%).
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI
Benchmarks: 3DMatch 3DLoMatch KITTI 3DMatch-EOR
Metrics: registration recall rotation error translation error inlier precision inlier number ratio inlier number feature matching recall
Baselines: PointDSC VBReg Hunter SM TEASER RANSAC-1M SC2-PCR MAC FastMAC
Builds on: SC2-PCR MAC RANSAC TEASER
Closest papers by meaning
- Scalable 3D Registration via Truncated Entry-wise Absolute Residuals (CVPR 2024 · similarity 0.88)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.86)
- Locality Preserving Refinement for Shape Matching with Functional Maps (AAAI 2024 · similarity 0.85)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.85)
- Doppelgangers++: Improved Visual Disambiguation with Geometric 3D Features (CVPR 2025 · similarity 0.84)
- Stable-SCore: A Stable Registration-based Framework for 3D Shape Correspondence (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.