Research map / Point clouds / Registration and matching
HybridReg: Robust 3D Point Cloud Registration with Hybrid Motions
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.
Builds HybridMatch/HybridLoMatch, a synthetic indoor point cloud registration dataset with rigid and non-rigid foreground motions, and proposes HybridReg, which learns Laplace-distributed uncertainty masks trained with an NLL loss to suppress deforming outliers in correspondence matching.
Problem
Scene-level 3D point cloud registration degrades when dynamic non-rigid foreground objects create outlier correspondences, and existing indoor datasets assume rigid motion while non-rigid datasets are object-level.
Approach
Reformulate correspondence estimation probabilistically: predict per-point uncertainty masks via cross-attention features, train them with a Laplace-based negative log-likelihood loss, and use (1 - uncertainty) to weight the Gaussian correlation matrix so unreliable foreground matches are suppressed.
Key results
- On HybridLoMatch 30-50% non-rigid, RR reaches 73.6% versus 61.0% for the second-best GeoTransformer (reported as 20.7% relative gain).
- On 3DMatch, RR is 92.7% (best) versus 92.0% for RegTR; on 3DLoMatch, RR is 73.1% vs 74.0% for GeoTransformer.
- On ETH with LGR, registration recall is 86.5% vs 85.6% for GeoTransformer+LGR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with non-rigid foreground motion data (ablation) | 3D-FRONT baseline · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 52.8 -> 90.4; 44.8 -> 86.2; 18.8 -> 54.0; 14.8 -> 48.6 | Helped |
| Add rigid foreground motion (ablation) | Hybrid-motion training data · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 90.4 -> 94.0; 86.2 -> 89.2; 54.0 -> 62.8; 48.6 -> 54.2 | Helped |
| Delete large flat planes (ablation) | Hybrid-motion training data · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 94.0 -> 97.4; 89.2 -> 96.6; 62.8 -> 75.6; 54.2 -> 73.6 | Helped |
| Add uncertainty-mask learning supervised with L1 loss (ablation) | GeoTrans · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 95.6 -> 94.2; 93.6 -> 95.2; 64.2 -> 66.2; 61.0 -> 61.8 | Mixed |
| Use NLL rather than L1 loss to supervise uncertainty-mask learning (ablation) | HybridReg with uncertainty mask · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 94.2 -> 97.2; 95.2 -> 96.6; 66.2 -> 75.5; 61.8 -> 73.2 | Helped |
| Use the predicted uncertainty mask to refine the Gaussian correlation matrix (ablation) | HybridReg with uncertainty mask and NLL loss · HybridMatch and HybridLoMatch, 10-30% and 30-50% non-rigid | RR: 97.2 -> 97.4; 96.6 -> 96.6; 75.5 -> 75.6; 73.2 -> 73.6 | Helped |
| Use HybridReg instead of GeoTrans | 3DMatch · 3DMatch | RR: 91.5 -> 92.7 | Helped |
| Use HybridReg instead of GeoTrans | 3DLoMatch · 3DLoMatch | RR: 74.0 -> 73.1 | Hurt |
| Use HybridReg + LGR instead of GeoTrans + LGR | Models trained on 3DMatch · ETH, 1000 samples | RR: 85.6 -> 86.5 | Helped |
Surprising findings
- Adding rigid foreground motion alone gives smaller gains than non-rigid motion, and models trained only on rigid data degrade sharply under hybrid motion, especially at low overlap.
Limitations
- Dataset is synthetic, so transfer to real captured scenes is not validated on hybrid motion.
- Gains on 3DLoMatch are mixed; on 3DMatch/3DLoMatch the method is not uniformly best.
Experimental setup
Datasets: 3D-FRONT ShapeNet DeformingThings4D 3DMatch ETH
Benchmarks: HybridMatch HybridLoMatch 3DMatch 3DLoMatch ETH
Metrics: Registration Recall RRE RTE Inlier Ratio Feature Matching Recall
Baselines: PREDATOR CoFiNet GeoTransformer PerfectMatch FCGF D3Feat RegTR
Builds on: GeoTransformer KPConv-FPN overlap-aware circle loss LGR PREDATOR
Closest papers by meaning
- Register Anything Model for Generalizable and Robust Point Cloud Registration (NeurIPS 2026 · similarity 0.88)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.88)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 2025 · similarity 0.87)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.86)
- Partial Point Cloud Registration with Multi-view 2D Image Learning (AAAI 2025 · similarity 0.86)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
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.