Research map / Point clouds / Point cloud classification
PaRot: Patch-Wise Rotation-Invariant Network via Feature Disentanglement and Pose Restoration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add intra-scale learning without relative pose restoration (ablation) | local-only PaRot · ModelNet40 z/SO3 | accuracy: 84.3 -> 87.4 | Helped |
| Add inter-scale learning without geometric relation representation (ablation) | local-only PaRot · ModelNet40 z/SO3 | accuracy: 84.3 -> 70.9 | Hurt |
| Add intra- and inter-scale learning without pose restoration (ablation) | local-only PaRot · ModelNet40 z/SO3 | accuracy: 84.3 -> 58.6 | Hurt |
| Add full pose restoration to intra-scale learning (ablation) | local + intra-scale PaRot · ModelNet40 z/SO3 | accuracy: 87.4 -> 89.5 | Helped |
| Add full pose restoration to inter-scale learning (ablation) | local + inter-scale PaRot · ModelNet40 z/SO3 | accuracy: 70.9 -> 89.6 | Helped |
| Add full pose restoration to intra- and inter-scale learning (ablation) | local + intra- and inter-scale PaRot · ModelNet40 z/SO3 | accuracy: 58.6 -> 91.0 | Helped |
| Use full relative orientation and position information instead of either part alone (ablation) | PaRot pose restoration · ModelNet40 z/SO3 | accuracy: 90.4, 90.5 -> 91.0 | Helped |
| Replace distance-weighted interpolation with pose-aware geometric-relation propagation (ablation) | ShapeNetPart segmentation · ShapeNetPart z/SO3, 3 neighbors | C.mIoU: 76.6 -> 77.6 | Helped |
| Increase pose-aware propagation neighbors from 3 to 13 (ablation) | PaRot segmentation · ShapeNetPart z/SO3 | I.mIoU: 82.1 -> 83.2 | Helped |
| Increase interpolation neighbors from 3 to 13 (ablation) | PointNet++ interpolation · ShapeNetPart z/SO3 | I.mIoU: 81.1 -> 80.6 | Hurt |
| Remove auxiliary siamese modules during testing (ablation) | PaRot · ModelNet40 z/SO3 | accuracy: 91.0 -> 91.0 | No clear effect |
Closest papers by meaning
- The Devil Is in the Pose: Ambiguity-Free 3D Rotation-Invariant Learning via Pose-Aware Convolution (CVPR 2022 · similarity 0.92)
- Rethinking Rotation Invariance with Point Cloud Registration (AAAI 2023 · similarity 0.89)
- Enhancing Rotation-Invariant 3D Learning with Global Pose Awareness and Attention Mechanisms (AAAI 2026 · similarity 0.89)
- Local-consistent Transformation Learning for Rotation-invariant Point Cloud Analysis (CVPR 2024 · similarity 0.89)
- Learning 3D Equivariant Implicit Function with Patch-Level Pose-Invariant Representation (NeurIPS 2024 · similarity 0.87)
- Leveraging 3D Geometric Priors in 2D Rotation Symmetry Detection (CVPR 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.