Research map / Point clouds / Registration and matching
Dynamic Cues-Assisted Transformer for Robust 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 DCATr, a transformer for keypoint-free point cloud registration that restricts cross-attention to non-local "cues" of potential correspondences and updates them dynamically, reaching state-of-the-art on 3DMatch, 3DLoMatch, and KITTI.
Problem
Ambiguous superpoint matching in low-overlap point cloud registration, where vanilla global cross-attention aggregates interference from irrelevant or inconsistent regions, causing wrong correspondences.
Approach
Extract non-local cues (top-k geometrically correlated superpoints) for each potential correspondence, restrict attention to these cues to model inter-frame consistency with confidence-weighted aggregation, and repeat this process dynamically several times over globally re-estimated correspondences.
Key results
- On 3DLoMatch with RANSAC-50K, DCATr reaches 76.8% registration recall at 5000 samples and inlier ratio up to 62.1% at 250 samples, outperforming GeoTrans by 2.5% to 11.7% in IR.
- With the LGR estimator on 3DLoMatch, DCATr achieves 75.7% RR and 48.2% IR versus GeoTrans 74.0% and 43.3%.
- On KITTI odometry, DCATr achieves RRE 0.22 deg and RTE 6.6 cm with RR 99.7%, lowest errors among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace GeoTrans with DCATr | 3DMatch registration · 3DMatch | IR (%): 70.3 -> 75.0 | Helped |
| replace GeoTrans with DCATr | 3DLoMatch registration · 3DLoMatch | RR (%): 74.0 -> 75.7 | Helped |
| replace GeoTrans with DCATr | KITTI odometry registration · KITTI odometry | RR (%): 99.8 -> 99.7 | Hurt |
| add dynamic updating with residual connection (ablation) | DCATr · 3DMatch | RR (%): 91.0 -> 92.1 | Helped |
| add dynamic updating with residual connection (ablation) | DCATr · 3DMatch | IR (%): 70.9 -> 75.0 | Helped |
| add dynamic updating with residual connection (ablation) | DCATr · 3DLoMatch | RR (%): 74.2 -> 75.7 | Helped |
| add dynamic updating with residual connection (ablation) | DCATr · 3DLoMatch | IR (%): 43.9 -> 48.2 | Helped |
| use a serial connection instead of residual connection (ablation) | static DCATr · 3DMatch | IR (%): 70.9 -> 66.8 | Hurt |
| use a serial connection instead of residual connection (ablation) | static DCATr · 3DLoMatch | IR (%): 43.9 -> 41.3 | Hurt |
Surprising findings
- A static residual variant of cues-assisted attention yields only slight gains over GeoTrans, while the dynamic updating is needed for larger improvement.
- Serial connection of the cues-assisted attention after prior extraction lowers performance below GeoTrans.
Limitations
- Relies on the coarse-to-fine pipeline and pose estimators (RANSAC, LGR) and gains on FMR are comparable rather than superior.
- Cue extraction and repeated dynamic updates add computation, with no efficiency analysis reported.
Experimental setup
Datasets: 3DMatch 3DLoMatch KITTI odometry
Benchmarks: 3DMatch 3DLoMatch KITTI odometry
Metrics: Inlier Ratio Registration Recall Feature Matching Recall Relative Rotation Error Relative Translation Error
Baselines: GeoTrans CoFiNet Predator D3Feat FCGF SpinNet YOHO 3DSNet SC2-PCR DGR HRegNet FMR
Builds on: GeoTrans CoFiNet KPConv-FPN Predator
Closest papers by meaning
- Dual Focus-Attention Transformer for Robust Point Cloud Registration (CVPR 2025 · similarity 0.90)
- A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration (NeurIPS 2024 · similarity 0.89)
- Multi-scale Consistency for Robust 3D Registration via Hierarchical Sinkhorn Tree (NeurIPS 2024 · similarity 0.88)
- Learning Instance-Aware Correspondences for Robust Multi-Instance Point Cloud Registration in Cluttered Scenes (CVPR 2024 · similarity 0.87)
- 3D Focusing-and-Matching Network for Multi-Instance Point Cloud Registration (NeurIPS 2024 · similarity 0.86)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 2025 · 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.