Research map / Point clouds / Point cloud classification
Local-consistent Transformation Learning for Rotation-invariant Point Cloud Analysis
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 LocoTrans, a two-branch DGCNN-based network that builds a local-consistent reference frame (LCRF) and a relative pose recovery (RPR) module for rotation-invariant point cloud classification and part segmentation.
Problem
Local reference frame (LRF)-based rotation-invariant point cloud analysis perturbs local geometric relations because learned LRF axes are inconsistent across neighbouring points, and relative pose between patches is lost in intermediate layers.
Approach
Constructs a local-consistent reference frame by choosing two symmetric axes around the angular bisector of equivariant features, enforces orthogonality and cross-point consistency losses, and restores relative pose between patches using equivariant-branch features in a relative pose recovery module.
Key results
- Achieves 91.6% accuracy on ModelNet40 under z/SO(3), the best among compared methods (PaRI-Conv 91.4%).
- Reaches 85.0% on ScanObjectNN OBJ_BG under z/SO(3), versus 84.3% for Li et al.
- Obtains 84.0% instance mIoU on ShapeNetPart under z/SO(3) and boosts DGCNN class mIoU by 42.7% in the z/SO(3) setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LocoTrans | DGCNN · ModelNet40 z/SO(3) | accuracy (%): 20.6 -> 91.6 | Helped |
| add LocoTrans | DGCNN · ScanObjectNN z/SO(3) | accuracy (%): 17.7 -> 85.0 | Helped |
| add LocoTrans | DGCNN · ShapeNetPart z/SO(3) | Insta. mIOU (%) | Helped |
| add fusion of invariant- and equivariant-branch outputs (ablation) | Invariant-branch baseline · ModelNet40 z/SO(3) | accuracy (%): 89.1 -> 90.4 | Helped |
| replace Gram-Schmidt LRF with LCRF (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 90.4 -> 90.7 | Helped |
| add RPR using point coordinates (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 90.4 -> 90.8 | Helped |
| add RPR using equivariant features (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 90.4 -> 91.3 | Helped |
| construct LRF with LCRF rather than hand-crafted/global-center and local-barycenter LRF (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 91.2 -> 91.6 | Helped |
| construct LRF with LCRF rather than hand-crafted/global-center and local-barycenter LRF (ablation) | LocoTrans · ScanObjectNN z/SO(3) | accuracy (%): 82.8 -> 85.0 | Helped |
| use equivariant features instead of coordinates for RPR (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 91.0 -> 91.6 | Helped |
| use equivariant features instead of invariant features for RPR (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 90.8 -> 91.6 | Helped |
| use hand-crafted APPF-w/oDirection instead of coordinates for RPR (ablation) | LocoTrans · ModelNet40 z/SO(3) | accuracy (%): 91.0 -> 91.0 | No clear effect |
Surprising findings
- Hand-crafted LRF based on global center and local barycenter underperforms the learned Gram-Schmidt LRF on ScanObjectNN, and the gap widens with background noise.
Limitations
- Two-branch design with equivariant branch increases parameters and FLOPs substantially.
- LCRF consistency only holds locally, so relative pose is still lost in intermediate layers, requiring RPR.
- Performance gap between z/SO(3) and SO(3)/SO(3) on ScanObjectNN (0.5%) attributed to learning-based LCRF sensitivity to augmentation randomness.
Experimental setup
Datasets: ModelNet40 ScanObjectNN ShapeNetPart
Benchmarks: ModelNet40 ScanObjectNN ShapeNetPart
Metrics: classification accuracy class mIoU instance mIoU
Baselines: PointNet PointNet++ DGCNN TFN VN-DGCNN SVNet-DGCNN RI-CNN RI-GCN AECNN SGMNet Li et al. PaRI-Conv
Builds on: DGCNN OrientedMP VN-DGCNN PaRI-Conv Gram-Schmidt orthogonalization
Closest papers by meaning
- 4D Local Modeling Toward Dynamic Global Perception for Ambiguity-free Rotation-Invariant Point Cloud Analysis (CVPR 2026 · similarity 0.89)
- Rethinking Rotation Invariance with Point Cloud Registration (AAAI 2023 · similarity 0.89)
- Rectified Point Flow: Generic Point Cloud Pose Estimation (NeurIPS 2025 · similarity 0.86)
- Learning SO(3)-Invariant Semantic Correspondence via Local Shape Transform (CVPR 2024 · similarity 0.86)
- Occlusion-aware Non-Rigid Point Cloud Registration via Unsupervised Neural Deformation Correntropy (ICLR 2025 · similarity 0.86)
- Rethinking Point Cloud Data Augmentation: Topologically Consistent Deformation (ICML 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.