Research map / Point clouds / Registration and matching
Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework
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 Cross-PCR, a feature-based cross-source point cloud registration framework combining a density-robust encoder with a loose-to-strict (one-to-many generation, consistency-based sparse and prior-guided dense) matching pipeline.
Problem
Cross-source point cloud registration fails because point clouds from different sensors (e.g., LiDAR vs. Kinect) differ greatly in density and distribution, leaving very few correct correspondences for feature-based registration.
Approach
A density-robust encoder fuses upsampled, original, and downsampled point features inside each encoder layer, and a "loose generation, strict selection" pipeline first generates many one-to-many correspondences then filters them via consistency-based sparse matching and globally prior-guided dense matching, followed by weighted SVD without RANSAC.
Key results
- On Kinect-LiDAR in 3DCSR, Cross-PCR reaches 66.7% RR, 35.5% IR and 73.2% FMR, improving RR by 57.6 pp over the prior best.
- On 3DMatch, Cross-PCR reports 94.5% RR and 88.7% IR, with IR gains of 6.1 pp over RoITr.
- On 3DMatch-DD with 0.1m voxel downsampling, Cross-PCR keeps 65.8% RR on 3DLoMatch-level setting while GeoTransformer drops by 73.3 pp.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the density-robust encoder instead of KPConv (ablation) | Cross-PCR · 3DCSR Kinect-lidar | RR: 57.7 -> 66.7 | Helped |
| use the density-robust encoder instead of KPConv (ablation) | Cross-PCR · 3DMatch-DD, voxel length 0.1 | RR: 57.6 -> 91.0 | Helped |
| add one-to-many loose correspondence generation instead of one-to-one matching (ablation) | Cross-PCR · 3DCSR Kinect-lidar | RR: 44.2 -> 66.7 | Helped |
| add consistency-based strict correspondence selection (ablation) | Cross-PCR · 3DCSR Kinect-lidar | RR: 27.3 -> 66.7 | Helped |
| add prior-guided global dense matching as a second stage (ablation) | Cross-PCR · 3DCSR Kinect-sfm | RR: 68.8 -> 93.8 | Helped |
| use Cross-PCR instead of GeoTransformer | 3DCSR Kinect-lidar · 3DCSR Kinect-lidar | RR: 9.1 -> 66.7 | Helped |
| use Cross-PCR instead of GeoTransformer | 3DMatch · 3DMatch, 5000 samples | RR: 92.0 -> 94.5 | Helped |
| use Cross-PCR instead of RoITr | 3DMatch · 3DMatch, 5000 samples | IR: 82.6 -> 88.7 | Helped |
| use Cross-PCR instead of RoITr | 3DMatch · 3DMatch, 5000 samples | FMR: 98.0 -> 97.9 | Hurt |
Surprising findings
- Strict selection is essential: removing it drops Kinect-LiDAR RR from 66.7% to 27.3%.
- Registration succeeds without RANSAC, using only weighted SVD on filtered correspondences.
Limitations
- Cross-source benchmark is limited to indoor scenes; authors note the lack of large cross-source datasets.
- FMR on 3DMatch is slightly lower than RoITr due to removal of ambiguous inliers.
- Some baselines were reproduced by the authors because code was unavailable.
Experimental setup
Datasets: 3DCSR 3DMatch
Benchmarks: 3DCSR 3DMatch 3DLoMatch 3DMatch-DD
Metrics: registration recall inlier ratio feature matching recall rotation error translation error
Baselines: GICP JRMPC GCTR GCC SpinNet Predator CoFiNet GeoTransformer YOHO RoITr
Builds on: KPConv GeoTransformer SC2-PCR circle loss TRIMs
Closest papers by meaning
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.90)
- SPEAL: Skeletal Prior Embedded Attention Learning for Cross-Source Point Cloud Registration (AAAI 2024 · similarity 0.88)
- Generative Point Cloud Registration (ICML 2025 · similarity 0.87)
- Zero-shot RGB-D Point Cloud Registration with Pre-trained Large Vision Model (CVPR 2025 · similarity 0.87)
- HeMoRa: Unsupervised Heuristic Consensus Sampling for Robust Point Cloud Registration (CVPR 2025 · similarity 0.86)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · 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.