Research map / Point clouds / Registration and matching
Multiway Point Cloud Mosaicking with Diffusion and Global Optimization
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 Wednesday, a multiway point cloud registration pipeline combining ODIN, a diffusion- and overlap-aware pairwise matcher, with rotation averaging, globally optimal translation re-estimation, and diffusion-based pose graph optimization.
Problem
Globally consistent alignment of many partially overlapping point clouds (multiway registration / point cloud mosaicking), where pairwise registration errors and low overlap cause failures in existing methods.
Approach
Integrates predicted point cloud overlap scores into attention learning with iterative refinement, denoises correlation matrices with a diffusion model, and adds a globally optimal consensus-maximization translation re-estimation step plus diffusion-based pose graph optimization adapted from prior work.
Key results
- ODIN reaches 69.73% registration recall on NSS versus 58.72% for PEAL, and 81.2% recall on 3DLoMatch versus 78.8% for PEAL.
- Wednesday with ODIN reduces NSS multiway rotation error to 2.01 degrees from 10.87 for the best prior method RMPR (roughly 80% reduction).
- The proposed globally optimal translation re-estimation runs in 0.05 s versus 0.12 s for exhaustive RANSAC, with more inliers (18.11 vs 18.03).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Reintegrate predicted overlap scores into attention learning (ablation) | ODIN · 3DLoMatch | RR (%): 0.791 -> 0.812 | Helped |
| Denoise the correlation matrix with diffusion (ablation) | ODIN · 3DLoMatch | RR (%): 0.803 -> 0.812 | Helped |
| Add overlap-score reintegration and correlation diffusion (ablation) | ODIN · 3DLoMatch | RR (%): 0.741 -> 0.812 | Helped |
| Use rotation and translation averaging rather than diffusion optimization alone (ablation) | Wednesday · NSS | RE (°): 5.21 -> 4.05 | Helped |
| Add translation re-estimation after rotation and translation averaging (ablation) | Wednesday · NSS | TE (m): 0.53 -> 0.48 | Helped |
| Add diffusion-based pose optimization to rotation and translation averaging (ablation) | Wednesday · NSS | RE (°): 4.05 -> 2.06 | Helped |
| Add diffusion-based pose optimization to rotation, translation re-estimation, and translation averaging (ablation) | Wednesday · NSS | TE (m): 0.48 -> 0.42 | Helped |
| Use ODIN instead of the selected baseline pairwise estimator | Wednesday · NSS | RE (°): 2.24 -> 2.01 | Helped |
Surprising findings
- Diffusion-based pose graph optimization alone yields the highest rotation error on NSS, requiring the other pipeline components to be effective.
Limitations
- Pairwise matcher is evaluated only in combination with multiple pipeline stages; ablations isolate components but not all stage interactions.
- Rotation averaging component is an older method (Chatterjee and Govindu) chosen because newer learning-based alternatives could not be reproduced.
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI NSS
Benchmarks: NSS 3DMatch 3DLoMatch KITTI
Metrics: Registration Recall (RR) Relative Rotation Error (RRE) Relative Translation Error (RTE) average rotation error (RE) average translation error (TE) runtime inlier number
Baselines: FPFH FCGF D3Feat RegTR Predator GeoTransformer PEAL Open3D global registration DeepMapping2 LMPR LIRTS RMPR
Builds on: GeoTransformer PEAL Diffusion-based pose graph optimization (Wang et al.) KPConv-FPN Overlap-aware circle loss
Closest papers by meaning
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.88)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.87)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- Where Precision Meets Efficiency: Transformation Diffusion Model for Point Cloud Registration (AAAI 2025 · similarity 0.87)
- Generative Point Cloud Registration (ICML 2025 · similarity 0.86)
- Partial Point Cloud Registration with Multi-view 2D Image Learning (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.