Research map / Point clouds / Registration and matching
3D Focusing-and-Matching Network for Multi-Instance 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 3DFMNet, a two-stage focusing-and-matching network that first regresses object centers via attention-based correlation with the CAD model and then performs dual-masked pairwise registration per object proposal for multi-instance point cloud registration.
Problem
Multi-instance point cloud registration: estimating poses of all instances of a model point cloud in a cluttered, occluded scene, where global correspondence plus clustering is unreliable.
Approach
Decompose one-to-many multi-instance registration into multiple pair-wise registrations by first localizing object centers (self/cross-attention correlation with the CAD model, offset and mask prediction, DBSCAN clustering) and then matching each ball-query object proposal to the model using an instance mask and an overlap mask.
Key results
- On ROBI, 3DFMNet reaches 46.81% MR, 50.61% MP and 48.63% MF, versus 38.51/41.19/39.80 for MIRETR.
- On Scan2CAD, 3DFMNet achieves 94.79% MF (94.15% MP), with MIRETR at 93.40% MF.
- Using ground-truth object centers as an upper bound gives 96.14% MF on Scan2CAD and 57.38% MF on ROBI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 3DFMNet instead of MIRETR | Multi-instance registration · ROBI test | MR (%): 38.51 -> 46.81 | Helped |
| Use 3DFMNet instead of MIRETR | Multi-instance registration · ROBI test | MP (%): 41.19 -> 50.61 | Helped |
| Use 3DFMNet instead of MIRETR | Multi-instance registration · ROBI test | MF (%): 39.80 -> 48.63 | Helped |
| Use 3DFMNet instead of ECC | Multi-instance registration · Scan2CAD test | MR (%): 96.52 -> 95.44 | Hurt |
| Use 3DFMNet instead of MIRETR | Multi-instance registration · Scan2CAD test | MP (%): 91.21 -> 94.15 | Helped |
| Use 3DFMNet instead of MIRETR | Multi-instance registration · Scan2CAD test | MF (%): 93.40 -> 94.79 | Helped |
| Use ground-truth object centers instead of predicted centers | 3DFMNet · Scan2CAD test | MF (%): 94.79 -> 96.14 | Helped |
| Use ground-truth object centers instead of predicted centers | 3DFMNet · ROBI test | MF (%): 48.63 -> 57.38 | Helped |
| Reduce sampled points from 4096 to 1024 (ablation) | 3DFMNet · Scan2CAD | MR (%): 95.44% -> 83.76% | Hurt |
| Increase sampled points from 4096 to 8192 (ablation) | 3DFMNet · Scan2CAD | MR (%): 95.44% -> 95.20% | Hurt |
| Add the overlap mask (ablation) | 3DFMNet · Scan2CAD | MR (%): 94.76% -> 95.44% | Helped |
| Add the overlap mask (ablation) | 3DFMNet · Scan2CAD | MP (%): 93.30% -> 94.15% | Helped |
Surprising findings
- On ROBI, the upper bound with ground-truth centers reaches only about 52% MR and 63% MP, showing that object-center localization is a main bottleneck.
Limitations
- Two-stage design means first-stage localization errors propagate to matching.
- Inference is slightly slower than the one-stage MIRETR.
- Upper bound on ROBI is low, indicating cluttered and incomplete objects remain hard.
Experimental setup
Datasets: Scan2CAD ROBI ShapeNet
Benchmarks: Scan2CAD ROBI
Metrics: MR MP MF RTE RRE PIR RMSE
Baselines: T-Linkage RansaCov PointCLM ECC MIRETR
Builds on: KPConv MIRETR PointCLM circle loss
Closest papers by meaning
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.92)
- Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes (CVPR 2024 · similarity 0.89)
- Dual Focus-Attention Transformer for Robust Point Cloud Registration (CVPR 2025 · similarity 0.88)
- Implicit Correspondence Learning for Image-to-Point Cloud Registration (CVPR 2025 · similarity 0.87)
- Multiway Point Cloud Mosaicking with Diffusion and Global Optimization (CVPR 2024 · similarity 0.87)
- Partial Point Cloud Registration with Multi-view 2D Image Learning (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.