pith. sign in

arxiv: 2211.14456 · v6 · submitted 2022-11-26 · 💻 cs.CV

TetraSphere: A Neural Descriptor for O(3)-Invariant Point Cloud Analysis

Pavlo Melnyk , Andreas Robinson , Michael Felsberg , M{\aa}rten Wadenb\"ack This is my paper

Pith reviewed 2026-05-24 11:00 UTC · model grok-4.3

classification 💻 cs.CV
keywords point cloud analysisO(3) invariancesteerable spherical neuronsvector neuronsrotation equivarianceTetraSphere3D deep learning
0
0 comments X

The pith

TetraSphere embeds steerable 3D spherical neurons into 4D vector neurons to create an O(3)-equivariant descriptor for point cloud analysis.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces TetraSphere, a neural network component that produces features unchanged under 3D rotations and reflections. It builds a TetraTransform that maps input point clouds from 3D to 4D space using steerable spherical neurons, then feeds the result into a vector-neuron version of DGCNN. The added module increases the total parameter count by less than 0.0002 percent. On randomly rotated real scans from ScanObjectNN the method reaches new state-of-the-art classification accuracy; on rotated synthetic ModelNet40 and ShapeNet data it surpasses prior equivariant approaches. Readers care because many practical 3D scans arrive in unknown orientations, and this construction supplies invariance without heavy augmentation or large extra computation.

Core claim

By constructing the TetraTransform as an equivariant 3D-to-4D embedding from steerable 3D spherical neurons and inserting it into the VN-DGCNN architecture, the resulting TetraSphere network extracts deeper O(3)-equivariant features and attains superior performance on O(3)-invariant point-cloud classification and segmentation tasks while adding fewer than 0.0002 percent extra parameters.

What carries the argument

TetraTransform: an equivariant embedding of 3D point-cloud input into 4D constructed from steerable 3D spherical neurons that enables subsequent vector-neuron processing.

If this is right

  • TetraSphere sets a new state-of-the-art on classification of randomly rotated real-world object scans from the challenging subsets of ScanObjectNN.
  • TetraSphere outperforms all prior equivariant methods on classification of randomly rotated objects from ModelNet40.
  • TetraSphere outperforms all prior equivariant methods on part segmentation of randomly rotated shapes from ShapeNet.
  • The integration of the TetraTransform into VN-DGCNN increases the parameter count by less than 0.0002 percent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The low parameter overhead suggests the same embedding could be inserted into other vector-neuron or equivariant backbones with little engineering cost.
  • If the spherical-neuron construction generalizes, similar low-cost embeddings might be derived for other symmetry groups such as SE(3) or for higher-dimensional inputs.
  • The reported stability on real-world rotated scans implies the method may reduce reliance on data-augmentation pipelines that multiply training time.

Load-bearing premise

The TetraTransform produces a faithful O(3)-equivariant embedding when placed inside the VN-DGCNN architecture without introducing representational loss or training instabilities.

What would settle it

Training TetraSphere on the rotated ScanObjectNN subsets and observing either lower accuracy than non-equivariant baselines that use rotation augmentation or repeated divergence during training would falsify the claim.

Figures

Figures reproduced from arXiv: 2211.14456 by Andreas Robinson, M{\aa}rten Wadenb\"ack, Michael Felsberg, Pavlo Melnyk.

Figure 1
Figure 1. Figure 1: Key component in our method (best viewed in color): a [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: High-level architecture of TetraSphere (for classification): the equivariant TT layer ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of the objects from the hardest subset of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Learned γ parameters for TetraSphereK=8 trained on the OBJ_BG subset of ScanObjectNN (see [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learned γ parameters for TetraSphereK=16 trained on the PB_T50_RS (see [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
read the original abstract

In many practical applications, 3D point cloud analysis requires rotation invariance. In this paper, we present a learnable descriptor invariant under 3D rotations and reflections, i.e., the O(3) actions, utilizing the recently introduced steerable 3D spherical neurons and vector neurons. Specifically, we propose an embedding of the 3D spherical neurons into 4D vector neurons, which leverages end-to-end training of the model. In our approach, we perform TetraTransform--an equivariant embedding of the 3D input into 4D, constructed from the steerable neurons--and extract deeper O(3)-equivariant features using vector neurons. This integration of the TetraTransform into the VN-DGCNN framework, termed TetraSphere, negligibly increases the number of parameters by less than 0.0002%. TetraSphere sets a new state-of-the-art performance classifying randomly rotated real-world object scans of the challenging subsets of ScanObjectNN. Additionally, TetraSphere outperforms all equivariant methods on randomly rotated synthetic data: classifying objects from ModelNet40 and segmenting parts of the ShapeNet shapes. Thus, our results reveal the practical value of steerable 3D spherical neurons for learning in 3D Euclidean space. The code is available at https://github.com/pavlo-melnyk/tetrasphere.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes TetraSphere, which embeds steerable 3D spherical neurons into 4D vector neurons via a TetraTransform step to produce an O(3)-equivariant descriptor. This is inserted into the VN-DGCNN architecture, yielding negligible parameter overhead (<0.0002%). The method is evaluated on randomly rotated point clouds and claims new state-of-the-art classification accuracy on challenging subsets of ScanObjectNN, as well as outperforming prior equivariant methods on ModelNet40 classification and ShapeNet part segmentation.

Significance. If the empirical results are confirmed, the work demonstrates that steerable spherical neurons can be practically combined with vector neurons to achieve strong rotation-invariant performance on real-world 3D data with almost no added cost. The public code release supports reproducibility and allows direct verification of the reported numbers.

major comments (2)
  1. [Method description (TetraTransform integration)] The central empirical claims rest on the TetraTransform producing a faithful O(3)-equivariant embedding inside VN-DGCNN without representational loss or training instabilities. The manuscript provides no explicit verification of this property (e.g., measured equivariance error, bijectivity check, or controlled ablation isolating the embedding step), which directly undermines attribution of the reported gains to the claimed invariance.
  2. [Experiments] Experimental results section: SOTA claims on rotated ScanObjectNN, ModelNet40, and ShapeNet lack reported statistical significance, number of independent runs, variance across seeds, and precise baseline re-implementations or hyper-parameter details. Without these, the performance margins cannot be assessed as robust.
minor comments (1)
  1. [Abstract] The abstract states the parameter increase is 'less than 0.0002%'; this figure should be derived explicitly from the added layers in the TetraTransform and reported with the exact total parameter counts for each model variant.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Method description (TetraTransform integration)] The central empirical claims rest on the TetraTransform producing a faithful O(3)-equivariant embedding inside VN-DGCNN without representational loss or training instabilities. The manuscript provides no explicit verification of this property (e.g., measured equivariance error, bijectivity check, or controlled ablation isolating the embedding step), which directly undermines attribution of the reported gains to the claimed invariance.

    Authors: We agree that explicit numerical verification of the equivariance property would strengthen attribution of the gains. In the revised manuscript we will add an appendix containing (i) a controlled ablation that isolates the TetraTransform step, (ii) measured equivariance error on held-out point clouds under random O(3) transformations, and (iii) a brief discussion of representational properties. These additions will be supported by the publicly released code. revision: yes

  2. Referee: [Experiments] Experimental results section: SOTA claims on rotated ScanObjectNN, ModelNet40, and ShapeNet lack reported statistical significance, number of independent runs, variance across seeds, and precise baseline re-implementations or hyper-parameter details. Without these, the performance margins cannot be assessed as robust.

    Authors: We acknowledge the value of reporting variance and statistical details. The revised version will include results averaged over five independent runs with different random seeds, reporting mean and standard deviation for all main tables. We will also expand the supplementary material with full hyper-parameter tables and explicit notes on how the baselines were re-implemented (using the original authors' code where available). revision: yes

Circularity Check

0 steps flagged

No significant circularity; equivariance by construction, performance empirical on external benchmarks

full rationale

The derivation constructs TetraTransform explicitly as an equivariant embedding from steerable 3D spherical neurons into 4D vector neurons, then inserts it into the existing VN-DGCNN architecture. Equivariance holds by the algebraic properties of the chosen components rather than by any fitted parameter or self-referential definition. All performance claims (SOTA on rotated ScanObjectNN subsets, ModelNet40, ShapeNet) are measured against independent external datasets and baselines, not recovered from the inputs by construction. No load-bearing self-citation, uniqueness theorem, or ansatz reduction appears in the chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the correctness of the O(3)-equivariant embedding produced by TetraTransform and its compatibility with vector-neuron layers; these are introduced without derivation from first principles and are validated only empirically.

axioms (1)
  • standard math O(3) group actions preserve the geometry of 3D Euclidean space
    Invariance under rotations and reflections is assumed as background group theory.
invented entities (2)
  • TetraTransform no independent evidence
    purpose: Equivariant embedding of 3D point clouds into 4D using steerable spherical neurons
    Newly proposed construction that maps 3D spherical neurons into 4D vector neurons.
  • TetraSphere no independent evidence
    purpose: Overall O(3)-invariant descriptor obtained by integrating TetraTransform into VN-DGCNN
    The named end-to-end model whose performance is claimed.

pith-pipeline@v0.9.0 · 5788 in / 1352 out tokens · 50431 ms · 2026-05-24T11:00:32.102347+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

57 extracted references · 57 canonical work pages · 3 internal anchors

  1. [1]

    Points to patches: Enabling the use of self-attention for 3d shape recognition

    Axel Berg, Magnus Oskarsson, and Mark O’Connor. Points to patches: Enabling the use of self-attention for 3d shape recognition. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 528–534. IEEE, 2022. 2

    work page 2022

  2. [2]

    Zz-net: A universal rotation equivariant architecture for 2d point clouds

    Georg Bökman, Fredrik Kahl, and Axel Flinth. Zz-net: A universal rotation equivariant architecture for 2d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10976–10985, 2022. 2

    work page 2022

  3. [3]

    ShapeNet: An Information-Rich 3D Model Repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information- rich 3d model repository. arXiv preprint arXiv:1512.03012,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis

    Chao Chen, Guanbin Li, Ruijia Xu, Tianshui Chen, Meng Wang, and Liang Lin. Clusternet: Deep hierarchical cluster network with rigorously rotation-invariant representation for point cloud analysis. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 4994–5002, 2019. 2

    work page 2019

  5. [5]

    Guibas, and He Wang

    Jiayi Chen, Yingda Yin, Tolga Birdal, Baoquan Chen, Leonidas J. Guibas, and He Wang. Projective manifold gra- dient layer for deep rotation regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6646–6655, 2022. 3

    work page 2022

  6. [6]

    The Devil is in the Pose: Ambiguity-free 3D Rotation-invariant Learning via Pose- aware Convolution

    Ronghan Chen and Yang Cong. The Devil is in the Pose: Ambiguity-free 3D Rotation-invariant Learning via Pose- aware Convolution. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 7472–7481, 2022. 1, 7, 8, 2

    work page 2022

  7. [7]

    Engineering applications of noncom- mutative harmonic analysis: with emphasis on rotation and motion groups

    Gregory S Chirikjian. Engineering applications of noncom- mutative harmonic analysis: with emphasis on rotation and motion groups. CRC press, 2000. 3

    work page 2000

  8. [8]

    3d-gfe: a three-dimensional geometric-feature extractor for point cloud data

    Yu-Chen Chou, Yen-Po Lin, Yang-Ming Yeh, and Yi-Chang Lu. 3d-gfe: a three-dimensional geometric-feature extractor for point cloud data. In 2021 Asia-Pacific Signal and Informa- tion Processing Association Annual Summit and Conference (APSIPA ASC), pages 2013–2017, 2021. 1, 2, 7

    work page 2021

  9. [9]

    Spherical CNNs

    Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    Vector neurons: A general framework for SO(3)-equivariant networks

    Congyue Deng, Or Litany, Yueqi Duan, Adrien Poulenard, Andrea Tagliasacchi, and Leonidas J Guibas. Vector neurons: A general framework for SO(3)-equivariant networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12200–12209, 2021. 1, 2, 3, 4, 5, 6, 7, 8

    work page 2021

  11. [11]

    Learning SO(3) Equivariant Represen- tations with Spherical CNNs

    Carlos Esteves, Christine Allen-Blanchette, Ameesh Makadia, and Kostas Daniilidis. Learning SO(3) Equivariant Represen- tations with Spherical CNNs. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2

    work page 2018

  12. [12]

    Rotpredictor: Unsupervised canonical viewpoint learning for point cloud classification

    Jin Fang, Dingfu Zhou, Xibin Song, Shengze Jin, Ruigang Yang, and Liangjun Zhang. Rotpredictor: Unsupervised canonical viewpoint learning for point cloud classification. In 2020 International Conference on 3D Vision (3DV), pages 987–996, 2020. 1, 3

    work page 2020

  13. [13]

    A versatile multi-view framework for lidar-based 3d object detection with guidance from panoptic segmentation

    Hamidreza Fazlali, Yixuan Xu, Yuan Ren, and Bingbing Liu. A versatile multi-view framework for lidar-based 3d object detection with guidance from panoptic segmentation. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17192–17201, 2022. 1

    work page 2022

  14. [14]

    SE(3)-Transformers: 3D Roto-Translation Equivari- ant Attention Networks

    Fabian Fuchs, Daniel Worrall, V olker Fischer, and Max Welling. SE(3)-Transformers: 3D Roto-Translation Equivari- ant Attention Networks. In Advances in Neural Information Processing Systems, pages 1970–1981. Curran Associates, Inc., 2020. 2

    work page 1970

  15. [15]

    Revisiting point cloud shape classification with a simple and effective baseline

    Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. International Conference on Machine Learning, 2021. 2

    work page 2021

  16. [16]

    Unrestricted recognition of 3d objects for robotics using multilevel triplet invariants

    Gosta H Granlund and Anders Moe. Unrestricted recognition of 3d objects for robotics using multilevel triplet invariants. AI Magazine, 25(2):51–51, 2004. 2

    work page 2004

  17. [17]

    Ng, and Zhiyong Wang

    Ruibin Gu, Qiuxia Wu, Hongbin Xu, Wing W.Y . Ng, and Zhiyong Wang. Learning efficient rotation representation for point cloud via local-global aggregation. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6, 2021. 2

    work page 2021

  18. [18]

    Ruibin Gu, Qiuxia Wu, Yuqiong Li, Wenxiong Kang, Wing W. Y . Ng, and Zhiyong Wang. Enhanced local and global learning for rotation-invariant point cloud representation. IEEE MultiMedia, 29(4):24–37, 2022. 1, 2

    work page 2022

  19. [19]

    Investigating the impact of multi-lidar place- ment on object detection for autonomous driving

    Hanjiang Hu, Zuxin Liu, Sharad Chitlangia, Akhil Agnihotri, and Ding Zhao. Investigating the impact of multi-lidar place- ment on object detection for autonomous driving. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2550–2559, 2022. 1

    work page 2022

  20. [20]

    Batch normalization: Accelerating deep network training by reducing internal co- variate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal co- variate shift. In International conference on machine learning, pages 448–456. PMLR, 2015. 4

    work page 2015

  21. [21]

    Rotation- invariant local-to-global representation learning for 3d point cloud

    Seohyun Kim, Jaeyoo Park, and Bohyung Han. Rotation- invariant local-to-global representation learning for 3d point cloud. Advances in Neural Information Processing Systems, 33:8174–8185, 2020. 3

    work page 2020

  22. [22]

    Deep projective 3d semantic segmentation

    Felix Järemo Lawin, Martin Danelljan, Patrik Tosteberg, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Deep projective 3d semantic segmentation. In International Conference on Computer Analysis of Images and Patterns , pages 95–107. Springer, 2017. 2

    work page 2017

  23. [23]

    A closer look at rotation-invariant deep point cloud analysis

    Feiran Li, Kent Fujiwara, Fumio Okura, and Yasuyuki Mat- sushita. A closer look at rotation-invariant deep point cloud analysis. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV) , pages 16218–16227,

    work page

  24. [24]

    A rotation-invariant framework for deep point cloud analysis

    Xianzhi Li, Ruihui Li, Guangyong Chen, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. A rotation-invariant framework for deep point cloud analysis. IEEE Transactions on Visualization and Computer Graphics, 2021. 1, 2

    work page 2021

  25. [25]

    PointCNN: Convolution on x-transformed points

    Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. PointCNN: Convolution on x-transformed points. Advances in neural information processing systems, 31:820–830, 2018. 2, 7

    work page 2018

  26. [26]

    Equivariant point cloud 9 analysis via learning orientations for message passing

    Shitong Luo, Jiahan Li, Jiaqi Guan, Yufeng Su, Chaoran Cheng, Jian Peng, and Jianzhu Ma. Equivariant point cloud 9 analysis via learning orientations for message passing. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18932–18941, 2022. 2

    work page 2022

  27. [27]

    Embed Me if You Can: A Geometric Perceptron

    Pavlo Melnyk, Michael Felsberg, and Mårten Wadenbäck. Embed Me if You Can: A Geometric Perceptron. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1276–1284, 2021. 2, 3, 4, 1

    work page 2021

  28. [28]

    Steerable 3D Spherical Neurons

    Pavlo Melnyk, Michael Felsberg, and Mårten Wadenbäck. Steerable 3D Spherical Neurons. In Proceedings of the 39th International Conference on Machine Learning, pages 15330– 15339. PMLR, 2022. 1, 2, 3, 4, 7

    work page 2022

  29. [29]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, pages 8024–8035, 2019. 7

    work page 2019

  30. [30]

    Spherical decision surfaces using conformal modelling

    Christian Perwass, Vladimir Banarer, and Gerald Sommer. Spherical decision surfaces using conformal modelling. In Joint Pattern Recognition Symposium, pages 9–16. Springer,

    work page

  31. [31]

    Adrien Poulenard and Leonidas J. Guibas. A functional ap- proach to rotation equivariant non-linearities for tensor field networks. In 2021 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 13169–13178,

    work page 2021

  32. [32]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

    work page

  33. [33]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2

    work page 2017

  34. [34]

    Image-to-lidar self-supervised distillation for autonomous driving data

    Corentin Sautier, Gilles Puy, Spyros Gidaris, Alexandre Boulch, Andrei Bursuc, and Renaud Marlet. Image-to-lidar self-supervised distillation for autonomous driving data. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) , pages 9891–9901,

    work page

  35. [35]

    3d-rotation-equivariant quaternion neural net- works

    Wen Shen, Binbin Zhang, Shikun Huang, Zhihua Wei, and Quanshi Zhang. 3d-rotation-equivariant quaternion neural net- works. In European Conference on Computer Vision, pages 531–547. Springer, 2020. 2

    work page 2020

  36. [36]

    Learning to orient surfaces by self-supervised spherical cnns

    Riccardo Spezialetti, Federico Stella, Marlon Marcon, Lu- ciano Silva, Samuele Salti, and Luigi Di Stefano. Learning to orient surfaces by self-supervised spherical cnns. Advances in Neural information processing systems, 33:5381–5392, 2020. 3

    work page 2020

  37. [37]

    Canonical capsules: Self-supervised capsules in canonical pose

    Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey E Hinton, and Kwang Moo Yi. Canonical capsules: Self-supervised capsules in canonical pose. In Advances in Neural Information Processing Systems, pages 24993–25005. Curran Associates, Inc., 2021. 1, 3

    work page 2021

  38. [38]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3D point clouds. arXiv preprint arXiv:1802.08219, 2018. 1, 2

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  39. [39]

    Revisiting point cloud classification: A new benchmark dataset and classifica- tion model on real-world data

    Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classifica- tion model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 2, 6

    work page 2019

  40. [40]

    Vision and Lie’s approach to invariance.Image and vision computing, 13(4):259–277, 1995

    Luc Van Gool, Theo Moons, Eric Pauwels, and André Oost- erlinck. Vision and Lie’s approach to invariance.Image and vision computing, 13(4):259–277, 1995. 1

    work page 1995

  41. [41]

    Sarma, Michael M

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for Learning on Point Clouds. ACM Trans. Graph., 38(5), 2019. 3, 7, 2

    work page 2019

  42. [42]

    Dynamic graph cnn for learning on point clouds

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. Acm Transactions On Graphics (tog), 38(5):1–12, 2019. 1, 2, 4, 5

    work page 2019

  43. [43]

    3D steerable CNNs: Learn- ing rotationally equivariant features in volumetric data

    Maurice Weiler, Mario Geiger, Max Welling, Wouter Boomsma, and Taco S Cohen. 3D steerable CNNs: Learn- ing rotationally equivariant features in volumetric data. In Advances in Neural Information Processing Systems, pages 10381–10392, 2018. 1

    work page 2018

  44. [44]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin- guang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1912–1920, 2015. 2, 6

    work page 1912

  45. [45]

    Walk in the Cloud: Learning Curves for Point Clouds Shape Analysis

    Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the Cloud: Learning Curves for Point Clouds Shape Analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915–924, 2021. 2

    work page 2021

  46. [46]

    Endowing deep 3d models with rotation invariance based on principal component analysis

    Zelin Xiao, Hongxin Lin, Renjie Li, Lishuai Geng, Hongyang Chao, and Shengyong Ding. Endowing deep 3d models with rotation invariance based on principal component analysis. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6, 2020. 3

    work page 2020

  47. [47]

    SGMNet: Learning rotation-invariant point cloud rep- resentations via sorted Gram matrix

    Jianyun Xu, Xin Tang, Yushi Zhu, Jie Sun, and Shiliang Pu. SGMNet: Learning rotation-invariant point cloud rep- resentations via sorted Gram matrix. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 10468–10477, 2021. 1, 2, 6

    work page 2021

  48. [48]

    Rethinking rotation invariance with point cloud registration

    Jianhui Yu, Chaoyi Zhang, and Weidong Cai. Rethinking rotation invariance with point cloud registration. In Proceed- ings of the AAAI Conference on Artificial Intelligence, pages 3313–3321, 2023. 3, 6, 7, 8, 1, 2

    work page 2023

  49. [49]

    Learning rotation-invariant rep- resentations of point clouds using aligned edge convolutional neural networks

    Junming Zhang, Ming-Yuan Yu, Ram Vasudevan, and Matthew Johnson-Roberson. Learning rotation-invariant rep- resentations of point clouds using aligned edge convolutional neural networks. In 2020 International Conference on 3D Vision (3DV), pages 200–209. IEEE, 2020. 2

    work page 2020

  50. [50]

    Rotation invariant convolutions for 3d point clouds deep learning

    Zhiyuan Zhang, Binh-Son Hua, David W Rosen, and Sai-Kit Yeung. Rotation invariant convolutions for 3d point clouds deep learning. In 2019 International Conference on 3D Vision (3DV), pages 204–213. IEEE, 2019. 2

    work page 2019

  51. [51]

    Rotation invariant point cloud classification: 10 Where local geometry meets global topology

    Chen Zhao, Jiaqi Yang, Xin Xiong, Angfan Zhu, Zhiguo Cao, and Xin Li. Rotation invariant point cloud classification: 10 Where local geometry meets global topology. arXiv preprint arXiv:1911.00195, 2019. 2, 6

    work page arXiv 1911

  52. [52]

    Quater- nion equivariant capsule networks for 3d point clouds

    Yongheng Zhao, Tolga Birdal, Jan Eric Lenssen, Emanuele Menegatti, Leonidas Guibas, and Federico Tombari. Quater- nion equivariant capsule networks for 3d point clouds. In Eu- ropean Conference on Computer Vision, pages 1–19. Springer,

    work page

  53. [53]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745–5753,

    work page

  54. [54]

    c 𝑹!#c c 𝐱 −𝟐𝑿

    3 11 TetraSphere: A Neural Descriptor for O(3)-Invariant Point Cloud Analysis Supplementary Material 𝑹!!c 𝑹!"c 𝑹!#c c 𝐱 −𝟐𝑿"𝑹!#𝑺 −𝟐𝑿"𝑹!"𝑺 −𝟐𝑿"𝑹!!𝑺 𝑟 𝑟 𝑟𝑟𝟎 −𝟐𝑿"𝑺 𝒀 ∈ℝ#×% 𝒀& ∈ℝ#$×% 𝐖 ∈ℝ#$×# 𝑹𝑺= (𝑹𝐜, ()(‖𝐜‖𝟐−𝑟𝟐),1) ∈ℝ5𝑺 = (𝐜, ()(‖𝐜‖𝟐−𝑟𝟐),1) ∈ℝ5𝑿 = (𝐱,−1,−()‖𝐱‖𝟐) ∈ℝ5 B(𝑺)X = " ∈ℝ4 Figure 4. (Best viewed in color.) Top: Tetra-basis projection is the output of ...

    work page

  55. [55]

    Additional illustrations In order to help the reader to understand the main concepts of our approach, i.e., prior work (steerable) spherical neu- rons [28] and vector neurons [10], as well as 4D tetra-basis projections (see Figure 1 and Section 4.1), we provide illus- trations in Figure 4

    work page

  56. [56]

    Learned Tetra-selection In this section, we present the Tetra-selection discussed in Section 5.3. As we can see from Figures 5 and 6, TetraS- phere learns all but one γ parameter of the spherical deci- sion surface (see (5)), defining the steerable neuron(6), to be close to 0, effectively always selecting one tetra-basis (out of K) during inference. We at...

    work page

  57. [57]

    OurTetraSphere achieves the best performance among equivariant methods in both tasks, consistently outperforming VN-DGCNN

    Synthetic data results We present a complete comparison of the methods trained on synthetic data to perform classification and part segmentation in Tables 5 and 6, respectively. OurTetraSphere achieves the best performance among equivariant methods in both tasks, consistently outperforming VN-DGCNN. Only the two RI methods PaRINet [6] and Yu et al. [48] o...

    work page