Research map / Point clouds / Registration and matching
PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud 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 PSReg, a coarse-to-fine point cloud registration framework that adds a prior-guided sparse mixture-of-experts (PSMoE) routing module, using GeoTransformer-derived superpoint correspondence priors, to improve feature discriminability in overlapping regions.
Problem
Point cloud registration under partial overlap struggles because ambiguous structures (e.g., floors, walls) in overlapping regions produce feature ambiguity and many outlier matches.
Approach
Vanilla SMoE routing lacks guidance to separate overlapping from non-overlapping tokens; PSMoE adds prior embeddings (ordered coding of superpoint correspondences from a pretrained GeoTransformer) to the tokens before top-1 routing so that potentially matching tokens are dispatched to the same expert.
Key results
- Achieves 95.7% registration recall on 3DMatch and 79.3% on 3DLoMatch (5000 samples), the highest RR among compared methods.
- Improves inlier ratio on 3DMatch to 75.8% (5000 samples) versus 74.6% for PEAL.
- Ablation: vanilla SMoE gives negligible gain (RR 94.2 to 94.2 on 3DMatch), while ordered prior coding raises 3DMatch RR to 95.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PSReg instead of PEAL | point cloud registration · 3DMatch (5000 samples) | Registration Recall (%): 94.4 -> 95.7 | Helped |
| use PSReg instead of PEAL | point cloud registration · 3DLoMatch (5000 samples) | Registration Recall (%): 78.9 -> 79.3 | Helped |
| increase the prior overlap threshold from 0.0 to 0.3 (ablation) | PSReg · 3DMatch (LGR, dense correspondences) | Registration Recall (%): 95.0 -> 93.1 | Hurt |
| increase the prior overlap threshold from 0.0 to 0.3 (ablation) | PSReg · 3DLoMatch (LGR, dense correspondences) | Registration Recall (%): 78.5 -> 77.5 | Hurt |
| add SMoE to the baseline (ablation) | PSReg baseline · 3DMatch (LGR, dense correspondences) | Registration Recall (%): 94.2 -> 94.2 | No clear effect |
| add vanilla SMoE to the baseline (ablation) | PSReg baseline · 3DLoMatch (LGR, dense correspondences) | Registration Recall (%): 77.8 -> 78.1 | Helped |
| add PSMoE with binary coding (ablation) | PSReg baseline · 3DMatch (LGR, dense correspondences) | Registration Recall (%): 94.2 -> 94.6 | Helped |
| replace binary coding with ordered coding in PCE (ablation) | PSReg with PSMoE · 3DMatch (LGR, dense correspondences) | Registration Recall (%): 94.6 -> 95.0 | Helped |
| add PSMoE with binary coding (ablation) | PSReg baseline · 3DLoMatch (LGR, dense correspondences) | Registration Recall (%): 77.8 -> 79.2 | Helped |
| replace binary coding with ordered coding in PCE (ablation) | PSReg with PSMoE · 3DLoMatch (LGR, dense correspondences) | Registration Recall (%): 79.2 -> 78.5 | Hurt |
| use ordered coding instead of binary coding in PCE (ablation) | PSReg with PSMoE · 3DMatch (LGR, dense correspondences) | Inlier Ratio (%): 73.9 -> 74.7 | Helped |
| use ordered coding instead of binary coding in PCE (ablation) | PSReg with PSMoE · 3DLoMatch (LGR, dense correspondences) | Inlier Ratio (%): 49.5 -> 49.7 | Helped |
Surprising findings
- Vanilla SMoE added to the baseline gives negligible improvement in registration performance, and its routing shows no clear clustering of overlapping regions.
- FMR on 3DLoMatch is lower than some baselines (e.g., GeoTransformer 88.8 vs. PSReg 87.1 at 1000 samples), attributed to sensitivity to prior quality.
Limitations
- Performance depends on quality of the prior superpoint correspondences from a pretrained GeoTransformer; FMR drops on 3DLoMatch when priors are poor.
- FMR on 3DLoMatch does not beat some baselines such as GeoTransformer and SIRA-PCR.
Experimental setup
Datasets: 3DMatch ModelNet40
Benchmarks: 3DMatch 3DLoMatch ModelNet40 ModelLoNet
Metrics: Feature Matching Recall Inlier Ratio Registration Recall Relative Rotation Error Relative Translation Error Chamfer distance
Baselines: FCGF D3Feat SpinNet YOHO Predator CoFiNet GeoTransformer OIF-PCR PEAL SIRA-PCR DCP-v2 RPM-Net
Builds on: GeoTransformer PEAL sparse mixture of experts KPConv-FPN local-to-global registration
Closest papers by meaning
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.88)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.87)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- PosDiffNet: Positional Neural Diffusion for Point Cloud Registration in a Large Field of View with Perturbations (AAAI 2024 · similarity 0.87)
- HybridReg: Robust 3D Point Cloud Registration with Hybrid Motions (AAAI 2025 · similarity 0.87)
- A robust inlier identification algorithm for point cloud registration via $\mathbf{\ell_0}$-minimization (NeurIPS 2024 · 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.