Research map / Point clouds / Registration and matching
PointMBF: A Multi-scale Bidirectional Fusion Network for Unsupervised RGB-D Point Cloud Registration
Mingzhi Yuan, Kexue Fu, Zhihao Li, Yucong Meng, Manning Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace unidirectional late fusion with multi-scale bidirectional fusion | PointMBF vs LLT · ScanNet (trained on ScanNet) | translation accuracy at 5cm: 80.4 -> 83.9 | Improved |
| train on 3DMatch and test on unseen ScanNet | PointMBF · ScanNet generalization | translation accuracy at 5cm: 81.0 vs LLT 76.9 | Improved |
| use multi-scale bidirectional fusion instead of direct concatenation (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 78.5 -> 81.0 | Improved |
| use transformer fusion in high-level space instead of bidirectional fusion (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 81.0 -> 74.7 | Hurt |
| add residual design in fusion modules (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 80.3 -> 81.0 | Improved |
| treat depth as an extra input channel of a shared visual network (ablation) | UR&R-style visual branch · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 72.4 -> 64.1 | Hurt |
| add fusion in encoding stage (ablation) | PointMBF with decode and concat · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 80.6 -> 81.0 | Improved |
| add fusion in decoding stage (ablation) | PointMBF with encode and concat · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 80.5 -> 81.0 | Improved |
| add final concatenation to encode+decode bidirectional fusion (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 80.3 -> 81.0 | Improved |
| add geometric-to-visual fusion on top of V2G + CAT (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 80.3 -> 81.0 | Improved |
| add visual-to-geometric fusion on top of G2V + CAT (ablation) | PointMBF · ScanNet (trained on 3DMatch) | translation accuracy at 5cm: 79.2 -> 81.0 | Improved |
| use weaker single branches (visual worse than LLT in rotation, poor geometric branch) with fusion | PointMBF · ScanNet | registration accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- REGTR — REGTR: End-to-End Point Cloud Correspondences With Transformers (CVPR 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Improving RGB-D Point Cloud Registration by Learning Multi-Scale Local Linear Transformation (ECCV 2022 · similarity 0.92)
- Zero-shot RGB-D Point Cloud Registration with Pre-trained Large Vision Model (CVPR 2025 · similarity 0.89)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.88)
- Partial Point Cloud Registration with Multi-view 2D Image Learning (AAAI 2025 · similarity 0.88)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.87)
- Generative Point Cloud Registration (ICML 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.
