Research map / Point clouds / Registration and matching
PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques
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 PointMC, a multi-instance point cloud registration framework that filters correspondences with local spatial consistency, searches maximal cliques on a compatibility graph to generate pose hypotheses, and clusters the low-dimensional poses into instances.
Problem
Estimating multiple rigid transformations between a model point cloud and a scene containing several (possibly overlapping) instances of it, under heavy outlier correspondences; prior global-consistency and high-dimensional spectral clustering methods are ambiguous and slow.
Approach
Search maximal cliques on the correspondence compatibility graph (after local-spatial-consistency filtering) to produce many pose hypotheses, then cluster the low-dimensional pose vectors rather than high-dimensional correspondence features, selecting one best transformation per cluster.
Key results
- On ModelNet40, PointMC reaches 94.85% MR, 99.76% MP, 97.24% MF at 0.04s, versus PointCLM's 93.10% MR and 96.29% MF.
- On Scan2CAD, PointMC reaches 84.65% MR and 80.29% MF, versus PointCLM's 81.26% MR and 77.15% MF.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PointMC instead of the strongest baseline, PointCLM | Multi-instance point cloud registration · ModelNet40 | MR: 93.10 -> 94.85 | Helped |
| Use PointMC instead of the strongest baseline, PointCLM | Multi-instance point cloud registration · Scan2CAD | MR: 81.26 -> 84.65 | Helped |
| Replace global spatial consistency with local spatial consistency (ablation) | PointMC · ModelNet40 | MR: 93.31 -> 94.85 | Helped |
| Replace global spatial consistency with local spatial consistency (ablation) | PointMC · Scan2CAD | MR: 82.75 -> 84.65 | Helped |
| Use second-order instead of first-order compatibility measure (ablation) | PointMC · ModelNet40 | MR: 93.24 -> 94.85 | Helped |
| Use second-order instead of first-order compatibility measure (ablation) | PointMC · Scan2CAD | MR: 82.26 -> 84.65 | Helped |
| Use Chameleon instead of DBSCAN for transformation clustering (ablation) | PointMC · ModelNet40 | MR: 94.85 -> 95.11 | Helped |
| Use Chameleon instead of DBSCAN for transformation clustering (ablation) | PointMC · Scan2CAD | MR: 84.65 -> 86.53 | Helped |
| Use DBSCAN instead of Chameleon for transformation clustering (ablation) | PointMC · ModelNet40 | Runtime (s): 0.07 -> 0.04 | Helped |
| Use DBSCAN instead of Chameleon for transformation clustering (ablation) | PointMC · Scan2CAD | Runtime (s): 0.11 -> 0.07 | Helped |
Surprising findings
- Replacing the commonly used clustering on high-dimensional features with low-dimensional pose clustering is reported to be faster, with DBSCAN cutting runtime by about 36-43% versus Chameleon while giving slightly lower accuracy.
Limitations
- Maximal clique search and filtering is time-consuming.
- The pose clustering algorithm is sensitive to parameters, which may hurt generalization.
- Scan2CAD evaluation uses a fine-tuned Predator for correspondences, so results depend on that front end.
Experimental setup
Datasets: ModelNet40 Scan2CAD ScanNet ShapeNet
Benchmarks: ModelNet40 Scan2CAD
Metrics: mean recall mean precision mean F1 runtime
Baselines: T-linkage RansaCov CONSAC ECC PointCLM
Builds on: MAC (maximal clique registration) PointCLM SC2-PCR (second-order compatibility) Predator Bron-Kerbosch
Closest papers by meaning
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.92)
- Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes (CVPR 2024 · similarity 0.89)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 2025 · similarity 0.88)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.87)
- Multiway Point Cloud Mosaicking with Diffusion and Global Optimization (CVPR 2024 · similarity 0.87)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 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.