Research map / Point clouds / Point cloud classification
3D-Rotation-Equivariant Quaternion Neural Networks
Wen Shen, Binbin Zhang, Shikun Huang, Zhihua Wei, Quanshi Zhang
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 |
|---|---|---|---|
| revise PointNet++ into a rotation-equivariant quaternion network (REQNN) | PointNet++ · ModelNet40 (tested with arbitrary rotations) | accuracy: 23.57 -> 63.95 | Improved |
| revise DGCNN into a REQNN | DGCNN · ModelNet40 (tested with arbitrary rotations) | accuracy: 30.05 -> 83.03 | Improved |
| revise PointConv into a REQNN | PointConv · ModelNet40 (tested with arbitrary rotations) | accuracy: 21.93 -> 78.14 | Improved |
| revise PointNet++ into a REQNN | PointNet++ · 3D MNIST (tested with arbitrary rotations) | accuracy: 44.15 -> 68.99 | Improved |
| revise DGCNN into a REQNN | DGCNN · 3D MNIST (tested with arbitrary rotations) | accuracy: 45.37 -> 82.09 | Improved |
| revise PointConv into a REQNN | PointConv · 3D MNIST (tested with arbitrary rotations) | accuracy: 44.63 -> 78.59 | Improved |
| add z-axis rotation data augmentation to baseline training (ablation) | PointNet++, DGCNN, PointConv · ModelNet40 (tested with arbitrary rotations) | accuracy: 23.57 -> 26.43; 30.05 -> 31.34; 21.93 -> 23.72 | No clear effect |
| add z-axis rotation data augmentation to baseline training (ablation) | PointNet++, DGCNN, PointConv · 3D MNIST (tested with arbitrary rotations) | accuracy: 44.15 -> 51.16; 45.37 -> 49.25; 44.63 -> 50.95 | Improved |
| switch from standard DGCNN to REQNN when test rotations are introduced | DGCNN · ModelNet40 NR/NR vs NR/AR | accuracy: 92.90 -> 83.03 (NR/NR); 29.74 -> 83.03 (NR/AR) | Mixed |
| use REQNN instead of PRIN and QE-Capsule network under arbitrary test rotations | REQNN (DGCNN) · ModelNet40 NR/AR | accuracy: 68.85 -> 83.03; 74.07 -> 83.03 | Improved |
| revise original networks into REQNNs (computational cost) | PointNet++, DGCNN, PointConv · ModelNet40 complexity | FLOPs (G): 0.87 -> 2.51; 3.53 -> 8.24; 1.44 -> 4.22 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Point2Sequence — Point2Sequence: Learning the Shape Representation of 3D Point Clouds with an Attention-Based Sequence to Sequence Network (AAAI 2019) · better on 1 of 2 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
- Interpretable Rotation-Equivariant Quaternion Neural Networks for 3D Point Cloud Processing (TPAMI 2024 · similarity 0.97)
- Quaternion Equivariant Capsule Networks for 3D Point Clouds (ECCV 2020 · similarity 0.90)
- Asymptotically Stable Quaternion-valued Hopfield-structured Neural Network with Periodic Projection-based Supervised Learning Rules (NeurIPS 2025 · similarity 0.87)
- A Characterization Theorem for Equivariant Networks with Point-wise Activations (ICLR 2024 · similarity 0.84)
- An Intuitive Multi-Frequency Feature Representation for SO(3)-Equivariant Networks (ICLR 2024 · similarity 0.84)
- TetraSphere: A Neural Descriptor for O(3)-Invariant Point Cloud Analysis (CVPR 2024 · similarity 0.84)
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.
