Research map / Point clouds / Registration and matching
Mining and Transferring Feature-Geometry Coherence for Unsupervised 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 INTEGER, an unsupervised outdoor point cloud registration method that mines pseudo-labels by combining feature-space anchor clustering with geometric compatibility filtering, and transfers them to a student via anchor-based contrastive learning and mixed-density training.
Problem
Unsupervised outdoor LiDAR point cloud registration lacks reliable optimization objectives: prior methods use overly strong geometric assumptions or noisy pseudo-labels that ignore high-level contextual feature information, and struggle with low overlap and density variation at long range.
Approach
Latent inlier correspondences cluster around positive anchors (mean features of existing inliers) in feature space; FGCM iteratively mines inliers via anchor similarity and filters outliers with spatial compatibility (SC2-PCR), while ABCont uses teacher anchors as universal inlier/outlier targets for the student and MDS learns density invariance from sparse downsampled views.
Key results
- On KITTI, INTEGER achieves mRR 84.0 vs EYOC 83.2 among unsupervised methods, with RR 54.2 at distance [40,50) vs EYOC 52.3.
- On nuScenes, mRR 63.1 vs EYOC 61.7; in KITTI-to-nuScenes generalization, mRR 62.6 vs EYOC 55.3.
- Teacher inlier ratio at first epoch (tIR@1st Epoch) reaches 81.2 on KITTI, above w/o FGC (67.6) and w/o S.T.I (71.9).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use INTEGER instead of EYOC | Unsupervised registration · KITTI | mRR: 83.2 -> 84.0 | Helped |
| Use INTEGER instead of EYOC | Unsupervised registration · nuScenes | mRR: 61.7 -> 63.1 | Helped |
| Use INTEGER instead of EYOC | KITTI-trained model evaluated on nuScenes · KITTI → nuScenes | mRR: 55.3 -> 62.6 | Helped |
| Add ABCont (ablation) | INTEGER · KITTI | mRR: 83.5 -> 84.0 | Helped |
| Add Per-Batch Self-Adaption (ablation) | INTEGER · KITTI | mRR: 80.9 -> 84.0 | Helped |
| Add Feature-Geometry Clustering (ablation) | INTEGER · KITTI | mRR: 82.8 -> 84.0 | Helped |
| Add Mixed-Density Student (ablation) | INTEGER · KITTI | mRR: 82.7 -> 84.0 | Helped |
| Use synthetic teacher initialization (ablation) | INTEGER · KITTI | mRR: 83.7 -> 84.0 | Helped |
| Use synthetic teacher initialization (ablation) | INTEGER · KITTI | tIR@1st Epoch: 71.9 -> 81.2 | Helped |
Surprising findings
- Removing synthetic teacher initialization drops tIR at first epoch from 81.2 to 71.9, and removing per-batch self-adaptation drops it to 43.3.
Limitations
- Performance depends on teacher quality; a noisy teacher in distant scenarios can impede feature-geometry clustering.
- Pseudo-label mining is slower than existing methods due to the iterative FGCM procedure.
Experimental setup
Datasets: KITTI nuScenes
Benchmarks: KITTI nuScenes
Metrics: mRR RR RRE RTE inlier ratio
Baselines: FCGF FCGF+C Predator SpinNet D3Feat CoFiNet Geometric Transformer EYOC RIENet
Builds on: EYOC PointContrast SC2-PCR InfoNCE PointDSC
Closest papers by meaning
- SuP: Sub-cloud Driven Point Cloud Registration (CVPR 2026 · similarity 0.90)
- BUFFER: Balancing Accuracy, Efficiency, and Generalizability in Point Cloud Registration (CVPR 2023 · similarity 0.89)
- PointDSC: Robust Point Cloud Registration Using Deep Spatial Consistency (CVPR 2021 · similarity 0.89)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.89)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · similarity 0.88)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 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.