Research map / Point clouds / Registration and matching
SPEAL: Skeletal Prior Embedded Attention Learning for Cross-Source 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 SPEAL, a transformer-based point cloud registration method that injects unsupervised medial-axis skeleton points and skeletal features as geometric priors, and introduces the KITTI CrossSource cross-source dataset.
Problem
Point cloud registration across heterogeneous sources (e.g., LiDAR vs. reconstructed point clouds) and in unstructured scenes suffers from noise, outliers, and differences in scale and density, and current methods ignore intrinsic geometric topology.
Approach
Extract skeleton points via learned convex combinations of input points (approximating the medial axis transform), embed skeleton-aware distance and angular structure into geometric self- and cross-attention, and augment coarse correspondences with spectrally denoised skeletal correspondences.
Key results
- On KITTI CrossSource, SPEAL reaches RRE 1.41°, RTE 0.58 m, RR 97.3%, versus GeoTransformer at 1.87°, 0.63 m, 96.8%.
- On GermanyForest3D low-overlap (<=30%), SPEAL reaches 91.7% RR versus 88.1% for GeoTransformer.
- On KITTI Odometry, SPEAL achieves RRE 0.23° and RR 99.8%, matching GeoTransformer on RR and best on RRE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SPEAL instead of GeoTrans. | KITTI CrossSource registration · KITTI CrossSource | RR (%): 96.8 -> 97.3 | Helped |
| Use SPEAL instead of GeoTrans. | GermanyForest3D registration · GermanyForest3D, overlap ≤ 30% | RR (%): 88.1 -> 91.7 | Helped |
| Use skeleton-aware self-attention and vanilla cross-attention rather than both vanilla attention types (ablation) | SPEAL · GermanyForest3D | RR (%): 99.3 -> 97.4 | Hurt |
| Use skeleton-aware self-attention and skeleton-aware cross-attention rather than both vanilla attention types (ablation) | SPEAL · GermanyForest3D | RR (%): 99.3 -> 97.1 | Hurt |
| Remove the Correspondence Dual-Sampler (CDS) module (ablation) | SPEAL · GermanyForest3D | RR (%): 99.3 -> 96.3 | Hurt |
| Apply spectral denoising to superpoint correspondences only (ablation) | CDS · GermanyForest3D | RR (%): 98.1 -> 97.8 | Hurt |
| Apply spectral denoising to superpoint and skeletal correspondences (ablation) | CDS · GermanyForest3D | RR (%): 98.1 -> 98.1 | No clear effect |
| Apply spectral denoising to skeletal correspondences only (ablation) | CDS · GermanyForest3D | RR (%): 98.1 -> 99.3 | Helped |
Surprising findings
- Spectral denoising helps only the skeletal correspondences; applying it to superpoint correspondences degrades performance.
Limitations
- Skeleton extraction module is trained separately and gradients are blocked from the backbone, a design chosen for stability.
- Skeleton correspondences on non-overlap regions can introduce outliers, requiring spectral denoising.
Experimental setup
Datasets: KITTI Odometry KITTI CrossSource GermanyForest3D
Benchmarks: KITTI CrossSource KITTI Odometry GermanyForest3D
Metrics: RRE RTE RR inlier ratio
Baselines: RANSAC FGR FCGF DGR HRegNet CoFiNet Predator PCAM REGTR GeoTransformer
Builds on: Geometric Transformer KPConv-FPN Overlap-aware Circle Loss Local-to-Global Registration PEAL
Closest papers by meaning
- Cross-PCR: A Robust Cross-Source Point Cloud Registration Framework (AAAI 2025 · similarity 0.88)
- FreeReg: Image-to-Point Cloud Registration Leveraging Pretrained Diffusion Models and Monocular Depth Estimators (ICLR 2024 · similarity 0.88)
- Mining and Transferring Feature-Geometry Coherence for Unsupervised Point Cloud Registration (NeurIPS 2024 · similarity 0.87)
- PSReg: Prior-guided Sparse Mixture of Experts for Point Cloud Registration (AAAI 2025 · similarity 0.86)
- SMART-PC: Skeletal Model Adaptation for Robust Test-Time Training in Point Clouds (ICML 2025 · similarity 0.86)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · 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.