Research map / Point clouds / Registration and matching
A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical 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 CAST, a coarse-to-fine point cloud registration Transformer with consistency-aware spot-guided attention and a lightweight sparse-to-dense fine matching module that avoids optimal transport and RANSAC-style hypothesis selection.
Problem
Coarse matching in learned point cloud registration is sparse and loose because cross-attention attends to irrelevant similar regions, forcing expensive optimal transport or RANSAC for outlier rejection, which hurts efficiency and scalability for real-time odometry.
Approach
Guides cross-attention to "spots" built from geometrically consistent, confident neighbors (matching score times normalized compatibility-graph degree), samples salient nodes for self-attention from the compatibility graph, and uses sparse keypoint-to-patch matching with compatibility graph embedding followed by dense local-attention refinement.
Key results
- On KITTI, CAST achieves RTE 2.5 cm and RR 100.0%, a 60.3% RTE improvement over DiffusionPCR.
- On 3DMatch, CAST with RANSAC reaches 95.2% RR; on 3DLoMatch it reaches 75.1% RR at about 0.18 s, roughly 10x faster than PEAL and DiffusionPCR.
- On nuScenes, CAST achieves RTE 0.12 m, RRE 0.20°, and RR 99.9%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAST instead of DiffusionPCR | KITTI registration · KITTI | RTE (cm): 6.3 -> 2.5 | Helped |
| Use CAST instead of the strongest reported RRE baseline | KITTI registration · KITTI | RRE (°): 0.23 -> 0.27 | Hurt |
| Use CAST instead of Point-to-Plane ICP | nuScenes registration · nuScenes | RTE (m): 0.15 -> 0.12 | Helped |
| Use CAST instead of DiffusionPCR | 3DMatch registration · 3DMatch, 1000 samples | Registration recall (%): 94.5 -> 95.2 | Helped |
| Use CAST instead of DiffusionPCR | 3DLoMatch registration · 3DLoMatch, 1000 samples | Registration recall (%): 79.2 -> 75.1 | Hurt |
| Add spot-guided cross-attention and consistency-aware self-attention (ablation) | CAST coarse matching · 3DMatch | RR (%): 94.45 -> 96.01 | Helped |
| Add spot-guided cross-attention (ablation) | CAST coarse matching · 3DLoMatch | RR (%): 72.07 -> 74.57 | Helped |
| Add multi-scale feature fusion (ablation) | CAST coarse matching · 3DMatch | RR (%): 94.14 -> 96.01 | Helped |
| Add overlap prediction (ablation) | CAST coarse matching · 3DLoMatch | RR (%): 74.91 -> 76.59 | Helped |
| Add dense registration (ablation) | CAST fine matching · KITTI | RTE (cm): 3.13 -> 2.51 | Helped |
| Add learnable dense correspondences (ablation) | CAST fine matching · KITTI | RTE (cm): 2.85 -> 2.51 | Helped |
| Add keypoint detection (ablation) | CAST fine matching · KITTI | RTE (cm): 3.58 -> 2.51 | Helped |
Surprising findings
- Rotation error is slightly worse than some coarse-to-fine methods on KITTI, attributed to numerical issues in SVD-based pose estimation and the RRE metric's zero-setting when trace exceeds 3.
- On the ETH generalization test, point-wise methods including CAST show lower registration recall than patch-wise SpinNet and BUFFER.
Limitations
- Sparse keypoints make feature matching recall worse than other coarse-to-fine methods on 3DLoMatch in extremely low-overlap cases.
- Rotation error is affected by SVD numerical inaccuracies, so RRE may not reflect true performance.
- Best 3DLoMatch results rely on combining with RANSAC for robustness at low overlap.
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI odometry nuScenes ETH
Benchmarks: 3DMatch 3DLoMatch KITTI nuScenes ETH
Metrics: registration recall relative translation error relative rotation error inlier ratio feature matching recall patch inlier ratio patch matching recall runtime
Baselines: GeoTransformer OIF-Net PEAL DiffusionPCR Predator CoFiNet HRegNet DGR MAC FastMAC SC2-PCR PointDSC
Builds on: ASTR PointDSC GeoTransformer KPConv linear attention
Closest papers by meaning
- Dynamic Cues-Assisted Transformer for Robust Point Cloud Registration (CVPR 2024 · similarity 0.89)
- GraphI2P: Image-to-Point Cloud Registration with Exploring Pattern of Correspondence via Graph Learning (CVPR 2025 · similarity 0.88)
- Dual Focus-Attention Transformer for Robust Point Cloud Registration (CVPR 2025 · similarity 0.88)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.87)
- APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction (NeurIPS 2025 · similarity 0.87)
- 3D Focusing-and-Matching Network for Multi-Instance 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.