Research map / Point clouds / Registration and matching
Unlocking Generalization Power in LiDAR 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 UGP, a LiDAR point cloud registration framework that removes inter-frame cross-attention, adds progressive self-attention and BEV image features, to improve cross-distance and cross-dataset generalization.
Problem
LiDAR point cloud registration generalizes poorly across distances (cross-distance) and across datasets (e.g., 64-line KITTI vs 32-line nuScenes), as cross-attention-based coarse-to-fine methods degrade under density shifts.
Approach
Eliminate inter-frame cross-attention so superpoint matching relies on intra-frame self-attention; add a distance-masked progressive self-attention that expands the attention range across layers, and fuse BEV image features into superpoint features.
Key results
- Cross-distance on KITTI (train 10m, test 20-40m) achieves mRR 94.5%, and 82.0% RR at KITTI@40m, versus 88.3% mRR for BUFFER.
- Cross-distance on nuScenes achieves mRR 91.4% with 72.3% RR at nuScenes@40m.
- Cross-dataset nuScenes to KITTI achieves mRR 90.9%, versus 84.7% for BUFFER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| eliminate cross-attention (ablation) | CoFiNet, GeoTrans, and PARE · KITTI cross-distance generalization | mRR: 49.7 -> 58.8; 28.2 -> 87.7; 25.4 -> 47.3 | Mixed |
| eliminate cross-attention (ablation) | UGP · KITTI@40m | RR: 2.2 -> 66.2 | Helped |
| add progressive self-attention (ablation) | UGP with cross-attention elimination · KITTI@40m | RR: 66.2 -> 71.2 | Helped |
| add progressive self-attention (ablation) | UGP with cross-attention elimination · KITTI@30m | RR: 95.1 -> 96.2 | Helped |
| add BEV fusion (ablation) | UGP with cross-attention elimination · KITTI@40m | RR: 66.2 -> 74.8 | Helped |
| add BEV fusion (ablation) | UGP with cross-attention elimination · KITTI@30m | RR: 95.1 -> 95.7 | Helped |
| use progressive self-attention and BEV fusion together (ablation) | UGP with cross-attention elimination · KITTI@40m | RR: 66.2 -> 82.0 | Helped |
| replace LGR with RANSAC50K (ablation) | UGP · KITTI cross-distance | mRR: 94.5 -> 92.4 | Hurt |
| replace LGR with MAC (ablation) | UGP · KITTI cross-distance | mRR: 94.5 -> 94.9 | Helped |
| use UGP instead of BUFFER | LiDAR point cloud registration · KITTI@40m, trained on KITTI@10m | RR: 61.2 -> 82.0 | Helped |
| use UGP instead of BUFFER | LiDAR point cloud registration · nuScenes@40m, trained on nuScenes@10m | RR: 15.2 -> 72.3 | Helped |
| use UGP instead of BUFFER | LiDAR point cloud registration · KITTI@40m, trained on nuScenes@10m | RR: 53.2 -> 74.1 | Helped |
Surprising findings
- Removing cross-attention from existing coarse-to-fine methods (CoFiNet, GeoTransformer, PARE) markedly improves their cross-distance generalization.
Limitations
- Cross-distance advantage over BUFFER is not uniform: BUFFER is higher on KITTI@30m at some RRE/RTE thresholds.
- Performance on KITTI@40m and nuScenes@40m remains well below perfect recall (74.1% and 72.3%).
Experimental setup
Datasets: KITTI nuScenes
Benchmarks: KITTI nuScenes
Metrics: Registration Recall (RR) mean Registration Recall (mRR) RRE RTE Patch Inlier Ratio Inlier Ratio
Baselines: FCGF SpinNet Predator CoFiNet GeoTransformer BUFFER PARE GCL
Builds on: GeoTransformer KPConv ResNet LGR
Closest papers by meaning
- PointNetLK Revisited (CVPR 2021 · similarity 0.88)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.86)
- GeoNLF: Geometry guided Pose-Free Neural LiDAR Fields (NeurIPS 2024 · similarity 0.86)
- A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration (NeurIPS 2024 · similarity 0.86)
- LiSu: A Dataset and Method for LiDAR Surface Normal Estimation (CVPR 2025 · similarity 0.85)
- LiDAR-Net: A Real-scanned 3D Point Cloud Dataset for Indoor Scenes (CVPR 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.