Research map / Point clouds / Registration and matching
Extend Your Own Correspondences: Unsupervised Distant Point Cloud Registration by Progressive Distance Extension
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 EYOC, a fully unsupervised distant point cloud registration method that trains a 3D sparse-conv feature extractor via a labeler-student EMA scheme with progressively increasing frame intervals, using spatial filtering, SC2-PCR speculative registration and NN-search correspondence rediscovery.
Problem
Distant point cloud registration (LiDARs on vehicles 5-50 m apart) needs expensive pose labels and generalizes poorly to new data distributions; no unsupervised outdoor solution existed.
Approach
Progressively extend the frame interval between consecutive LiDAR sweeps so a labeler trained on near pairs labels farther pairs, using spatial filtering of close-to-LiDAR correspondences (which are less stable) and registration-based NN-search rediscovery to generate clean correspondence supervision without pose labels.
Key results
- EYOC reaches 83.2% mRR on KITTI when trained from scratch unsupervised, +11.4% over fixed supervised FCGF and only 4.7% below supervised Predator (87.9%).
- On out-of-domain nuScenes, EYOC finetuned from WOD achieves 68.4% mRR, +1.3% over FCGF.
- Time analysis: EYOC total training time 865 ms per step vs 1176 ms for FCGF (which needs two training routines).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply EYOC unsupervised training to FCGF on KITTI | FCGF · KITTI | mRR: 71.8 -> 83.2 | Helped |
| Apply EYOC unsupervised training to FCGF on WOD | FCGF · WOD | mRR: 69.9 -> 78.4 | Helped |
| Apply EYOC unsupervised training to FCGF on nuScenes | FCGF · nuScenes | mRR: 67.1 -> 61.7 | Hurt |
| Remove Lowe filtering (ablation) | EYOC · KITTI | labeler IR: 5.1 -> 18.4; 7.8 -> 43.3 | Helped |
| Add adaptive spatial filtering, speculative registration, and correspondence rediscovery (ablation) | EYOC · KITTI | mRR: 84.6 -> 88.0 | Helped |
| Add hard spatial filtering, speculative registration, and correspondence rediscovery (ablation) | EYOC · KITTI | RR @ [40,50]: 60.3 -> 68.8 | Helped |
| Remove the progressive dataset strategy (ablation) | EYOC · KITTI | labeler IR: 43.3 -> 0.0 | Hurt |
| Use EMA decay factor λ = 0.85 or 0.9 instead of λ = 0.2 (ablation) | EYOC · KITTI | RR @ [40,50]: 73.9 -> 58.8; 73.9 -> N/C | Hurt |
| Increase the hard spatial-filter threshold dthresh (ablation) | EYOC · KITTI | labeler IR: 18.4; 18.5; 25.2; 31.1; 31.4; 29.4; 45.1; 49.0; 53.2; 43.3 | Mixed |
| Increase the adaptive spatial-filter threshold sthresh (ablation) | EYOC · KITTI | labeler IR: 18.4; 25.4; 30.7; 31.5; 31.1; 34.9; 43.3; N/C; N/C; N/C | Mixed |
Surprising findings
- Lowe ratio filtering, effective for indoor RGB-D self-supervision, deteriorates labeler correspondence quality on outdoor LiDAR.
- Predator generalizes much worse than FCGF to out-of-domain nuScenes (34.5% vs 67.1% mRR).
Limitations
- Performance on nuScenes from scratch (61.7% mRR) is below FCGF generalization, and EYOC underperforms supervised FCGF on WOD by 10.9%.
- Sensitive to hyperparameters such as EMA decay and spatial filter thresholds; training diverges with inappropriate settings.
- Relies on consecutive LiDAR sweeps and approximate frame intervals as a proxy for distance, and needs a pretrained model or careful schedule to avoid divergence.
Experimental setup
Datasets: KITTI nuScenes Waymo Open Dataset (WOD)
Benchmarks: KITTI nuScenes WOD
Metrics: Registration Recall (RR) mean Registration Recall (mRR) Relative Rotation Error Relative Translation Error Inlier Ratio
Baselines: FCGF Predator SpinNet D3Feat CoFiNet GeoTransformer FCGF + C
Builds on: FCGF SC2-PCR Hardest-Contrastive Loss BYOC APR staged training
Closest papers by meaning
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework (AAAI 2025 · similarity 0.85)
- DV-Matcher: Deformation-based Non-rigid Point Cloud Matching Guided by Pre-trained Visual Features (CVPR 2025 · similarity 0.85)
- PointMC: Multi-instance Point Cloud Registration based on Maximal Cliques (ICML 2024 · similarity 0.85)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.85)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.85)
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.