Research map / Point clouds / Registration and matching
Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree
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 Hierarchical Sinkhorn Tree (HST), a BFS-pruned K-ary tree over feature-pyramid neighborhoods with an overlap-aware Sinkhorn distance, to filter coarse correspondences by multi-scale consistency for robust point cloud registration.
Problem
Coarse-to-fine point cloud registration suffers from noisy coarse correspondences under low overlap and high noise, and outlier-rejection methods struggle when outlier rates are high.
Approach
Models multi-scale consistency of each coarse correspondence by recursively exploring k-NN patches across feature-pyramid layers, scoring patch similarity with an overlap-aware Sinkhorn distance that only transports likely-overlapping points, and pruning the tree with the retained overlap points.
Key results
- HST reaches 93.5% RR on 3DMatch and 77.8% RR on 3DLoMatch with RANSAC-50k at 5000 samples, beating the prior best by 1.1 and 1.7 pp.
- With LGR, HST improves RR over GeoTransformer by 1.7% on 3DMatch (93.2%) and 3.6% on 3DLoMatch (77.3%).
- On KITTI odometry HST achieves RRE 0.23, RTE 6.3 cm, and RR 99.8%, the best on all metrics.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HST instead of OIF | point cloud registration · 3DMatch, 5000 samples | RR (%): 92.4 -> 93.5 | Helped |
| Use HST instead of OIF | point cloud registration · 3DLoMatch, 5000 samples | RR (%): 76.1 -> 77.8 | Helped |
| Add overlap-aware Sinkhorn Distance, including overlap filtering and overlap-aware initialization (ablation) | HST · 3DMatch, 5000 samples | RR (%): 90.4 -> 93.5 | Helped |
| Add overlap-aware Sinkhorn Distance, including overlap filtering and overlap-aware initialization (ablation) | HST · 3DLoMatch, 5000 samples | RR (%): 73.3 -> 77.8 | Helped |
| Add overlap filtering (ablation) | HST · 3DMatch, 5000 samples | RR (%): 91.6 -> 93.5 | Helped |
| Add overlap filtering (ablation) | HST · 3DLoMatch, 5000 samples | RR (%): 75.1 -> 77.8 | Helped |
| Add overlap-aware initialization (ablation) | HST · 3DMatch, 5000 samples | RR (%): 92.9 -> 93.5 | Helped |
| Add overlap-aware initialization (ablation) | HST · 3DLoMatch, 5000 samples | RR (%): 77.0 -> 77.8 | Helped |
| Add Patch Overlap Prediction (ablation) | HST · 3DMatch, 5000 samples | RR (%): 92.0 -> 93.5 | Helped |
| Add Patch Overlap Prediction (ablation) | HST · 3DLoMatch, 5000 samples | RR (%): 75.6 -> 77.8 | Helped |
| Increase HST depth from 1 to the full depth of 2 (ablation) | HST · 3DMatch, 5000 samples | RR (%): 92.8 -> 93.5 | Helped |
| Increase HST depth from 1 to the full depth of 2 (ablation) | HST · 3DLoMatch, 5000 samples | RR (%): 76.8 -> 77.8 | Helped |
Surprising findings
- Point-to-node grouping performs slightly worse than k-NN grouping for local exploration, and the authors note the difference is small.
Limitations
- Relies on quality of coarse correspondences; extreme low-overlap or noisy cases can still fail.
- Patch similarity is assessed purely geometrically, limited by noise and locality; color, normals or semantics are not used.
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI odometry 3DExMatch
Benchmarks: 3DMatch 3DLoMatch KITTI odometry 3DExMatch
Metrics: registration recall feature matching recall inlier ratio relative rotation error relative translation error
Baselines: FCGF D3Feat SpinNet Predator CoFiNet YOHO GeoTransformer RIGA REGTR OIF-PCR RoITr PEAL
Builds on: GeoTransformer KPConv-FPN Sinkhorn distances SuperGlue Predator
Closest papers by meaning
- Dynamic Cues-Assisted Transformer for Robust Point Cloud Registration (CVPR 2024 · similarity 0.88)
- A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.87)
- Scalable 3D Registration via Truncated Entry-wise Absolute Residuals (CVPR 2024 · similarity 0.87)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · similarity 0.87)
- Bridge 2D-3D: Uncertainty-aware Hierarchical Registration Network with Domain Alignment (AAAI 2025 · similarity 0.87)
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.