Research map / Point clouds / Point cloud classification
TetraSphere: A Neural Descriptor for O(3)-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 TetraSphere, an O(3)-invariant point cloud descriptor that lifts 3D points into 4D via steerable 3D spherical neurons (TetraTransform) and feeds them into a VN-DGCNN backbone.
Problem
Rotation- and reflection-invariant classification and part segmentation of 3D point clouds, especially for real-world scans under arbitrary rotations.
Approach
Embed steerable 3D spherical neurons as a learnable TetraTransform layer that lifts 3D points to O(3)-equivariant 4D vectors across K tetra-bases, max-pooled over K, then process with vector neurons and inner-product invariants, trained end-to-end.
Key results
- On ScanObjectNN PB_T50_RS (hardest, SO(3)/SO(3)), TetraSphere reaches 79.0% vs. 78.5% for VN-DGCNN, setting state of the art.
- On ScanObjectNN OBJ_BG under z/SO(3), TetraSphere K=2 reaches 87.3% (best 87.3 at K=2) vs. 83.5% for VN-DGCNN.
- TetraSphere adds fewer than 0.0002% parameters relative to VN-DGCNN (5 parameters per sphere); on ShapeNet part segmentation it reaches 82.3 mIoU vs. 81.4 for VN-DGCNN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the TetraTransform 4D representation to VN-DGCNN | VN-DGCNN · ScanObjectNN PB_T50_RS | accuracy (%): 78.5 -> 79.2 | Helped |
| increase the number of steerable neurons beyond K=2 (ablation) | TetraSphere · ScanObjectNN OBJ_BG | accuracy (%): 87.3; 84.5; 86.2; 85.4 | Mixed |
| append the point norm as a fourth input component (ablation) | VN-DGCNN · ScanObjectNN OBJ_BG | accuracy (%): 82.8 ± 0.6 (83.5) -> 82.7 ± 1.6 (84.5) | Hurt |
| add an initial equivariant VN layer (ablation) | VN-DGCNN · ScanObjectNN OBJ_BG | accuracy (%): 82.8 ± 0.6 (83.5) -> 83.3 ± 0.8 (83.8) | Helped |
| add an initial equivariant VN layer and append point norm (ablation) | VN-DGCNN · ScanObjectNN OBJ_BG | accuracy (%): 82.8 ± 0.6 (83.5) -> 82.7 ± 0.1 (82.8) | Hurt |
| apply O(3) transformations at test time, including reflections (ablation) | TetraSphere K=2 · ScanObjectNN OBJ_BG | accuracy (%): 85.5 ± 2.2 (87.3) -> 85.5 ± 2.2 (87.3) | No clear effect |
Surprising findings
- TetraSphere learns to select the same steerable neuron for all inputs, with all but one gamma parameter near zero, so the network could be pruned to K=1 at inference.
- Simply appending the norm to the 3D input ([x, ||x||]) to VN-DGCNN does not improve accuracy, indicating the gain is not from the extra invariant component alone.
Limitations
- Gains over VN-DGCNN on ScanObjectNN hardest subset are small (about 0.5 points) and differ across K, with no consistent trend in K.
- Increased inference time compared to VN-DGCNN due to 4D vector operations and TetraTransform.
- Evaluated only on classification and part segmentation with point cloud inputs; no features or normals.
Experimental setup
Datasets: ScanObjectNN ModelNet40 ShapeNet-part
Benchmarks: ScanObjectNN OBJ_BG ScanObjectNN PB_T50_RS ModelNet40 ShapeNet-part
Metrics: classification accuracy mIoU
Baselines: VN-DGCNN PaRINet Yu et al. (AIT) Li et al. (RI-Framework) 3D-GFE PointCNN DGCNN TFN
Builds on: VN-DGCNN vector neurons steerable 3D spherical neurons DGCNN
Closest papers by meaning
- Steerable 3D Spherical Neurons (ICML 2022 · similarity 0.89)
- Equivariant Point Network for 3D Point Cloud Analysis (CVPR 2021 · similarity 0.88)
- Rotation-Invariant Local-to-Global Representation Learning for 3D Point Cloud (NeurIPS 2020 · similarity 0.87)
- ZZ-Net: A Universal Rotation Equivariant Architecture for 2D Point Clouds (CVPR 2022 · similarity 0.86)
- SpinNet: Learning a General Surface Descriptor for 3D Point Cloud Registration (CVPR 2021 · similarity 0.86)
- Cross-Domain 3D Equivariant Image Embeddings (ICML 2019 · 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.