Research map / Point clouds / Registration and matching
Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes
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 MIRETR, a coarse-to-fine transformer that learns instance-aware superpoint features and instance masks to extract instance-wise correspondences for multi-instance point cloud registration, avoiding multi-model fitting.
Problem
Multi-instance point cloud registration in cluttered scenes, where global context pollutes superpoint features from background and other instances, and correspondences cannot be clustered reliably into instances.
Approach
An Instance-aware Geometric Transformer jointly predicts per-superpoint instance masks and uses them to restrict geometric context aggregation within instances, iteratively refining features and masks; instance candidates are then built from masked superpoint correspondences and merged by NMS-style selection instead of multi-model fitting.
Key results
- On ROBI, the full MIRETR pipeline reaches 39.80 MF versus 23.20 for the best GeoTransformer-based baseline (GeoTransformer + ECC), a gain of about 16.6 points.
- On Scan2CAD, MIRETR full pipeline achieves 93.40 MF and 95.70 IR, surpassing the best baseline GeoTransformer + ECC (83.57 MF, 94.63 IR).
- On ShapeNet unseen categories, MIRETR full pipeline reaches 94.44 MF versus 70.37 for GeoTransformer + ECC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instance-aware point matching (ablation) | MIRETR · ROBI | MF (%): 21.88 -> 39.80 | Helped |
| add the instance masking block (ablation) | MIRETR · ROBI | MR (%): 17.07 -> 41.19 | Helped |
| add instance-aware point matching and the instance masking block (ablation) | MIRETR · ROBI | MF (%): 18.65 -> 39.80 | Helped |
| replace global geometric self-attention with the geometric encoding block (ablation) | MIRETR · ROBI | MF (%): 10.96 -> 39.80 | Helped |
Surprising findings
- Replacing the geometric encoding block with global geometric self-attention gave the worst ablation result on ROBI (MF 10.96), indicating global context pollutes superpoint features.
- ECC-based model achieves slightly higher mean recall than full MIRETR on Scan2CAD but much lower precision, so high recall came from over-predicting registrations.
Limitations
- Relies on iterative mask refinement and thresholds (tau, tau_s), whose sensitivity is not discussed in the main text.
- Evaluated only on rigid, geometric-only settings with a single model per scene; multi-modal extension left to future work.
Experimental setup
Datasets: Scan2CAD ROBI ShapeNet
Benchmarks: Scan2CAD ROBI ShapeNet
Metrics: Mean Recall Mean Precision Mean F1 score Inlier Ratio instance mIoU
Baselines: FCGF CoFiNet GeoTransformer T-Linkage RansaCov ECC PointCLM
Builds on: GeoTransformer CoFiNet KPFCNN circle loss PointCLM
Closest papers by meaning
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.89)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.89)
- Dynamic Cues-Assisted Transformer for Robust Point Cloud Registration (CVPR 2024 · similarity 0.87)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 2025 · similarity 0.86)
- Dual Focus-Attention Transformer for Robust Point Cloud Registration (CVPR 2025 · similarity 0.86)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 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.