Research map / Point clouds / Registration and matching
PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration
Runzhao Yao, Shaoyi Du, Wenting Cui, Canhui Tang, Chengwu Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace vanilla VN convolution with PARE-Conv (node features) (ablation) | PARE-Net backbone · 3DLoMatch | Registration Recall (%): 76.6 -> 78.8 | Improved |
| replace vanilla VN convolution with PARE-Conv (edge features) (ablation) | PARE-Net backbone · 3DLoMatch | Registration Recall (%): 77.8 -> 80.5 | Improved |
| use feature-based hypothesis proposer instead of RANSAC (ablation) | PARE-Net pose estimation · 3DLoMatch | Registration Recall (%): 77.5 -> 80.5 | Improved |
| use feature-based hypothesis proposer instead of LGR (ablation) | PARE-Net pose estimation · 3DLoMatch | Registration Recall (%): 78.9 -> 80.5 | Improved |
| use feature-based hypothesis proposer instead of RANSAC (ablation) | PARE-Net pose estimation · 3DMatch | Registration Recall (%): 93.7 -> 95.0 | Improved |
| add contrastive rotation loss on equivariant features (ablation) | PARE-Net · 3DLoMatch | Registration Recall (%): 79.3 -> 80.5 | Improved |
| add contrastive rotation loss on equivariant features (ablation) | PARE-Net · 3DMatch | Registration Recall (%): 93.3 -> 95.0 | Improved |
| use L2 loss on equivariant features instead of no loss (ablation) | PARE-Net · 3DLoMatch | Registration Recall (%): 79.3 -> 75.4 | Hurt |
| use L2 loss on equivariant features instead of no loss (ablation) | PARE-Net · 3DMatch | Registration Recall (%): 93.3 -> 94.2 | Improved |
| rotation-equivariant design instead of rotation augmentation only (PARE-Net vs GeoTrans) | point cloud registration · Rotated 3DLoMatch | Transformation Recall (%): GeoTrans 75.4 -> 72.6 under rotation; Ours 81.3 -> 81.8 | Improved |
| PARE-Net vs PEAL in speed and recall | PARE-Net · 3DLoMatch | Time (s): 1.19 -> 0.17 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PEAL — PEAL: Prior-Embedded Explicit Attention Learning for Low-Overlap Point Cloud Registration (CVPR 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SHReg: Strictly Rotation-Equivariant Point Cloud Registration via Spherical Harmonics (ECCV 2026 · similarity 0.91)
- Equivariant Point Network for 3D Point Cloud Analysis (CVPR 2021 · similarity 0.90)
- Rethinking Rotation Invariance with Point Cloud Registration (AAAI 2023 · similarity 0.90)
- SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration (CVPR 2021 · similarity 0.90)
- PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration (AAAI 2023 · similarity 0.89)
- Local-consistent Transformation Learning for Rotation-invariant Point Cloud Analysis (CVPR 2024 · similarity 0.88)
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.
