pith. sign in

arxiv: 2605.15368 · v1 · pith:KXHWQWJRnew · submitted 2026-05-14 · 💻 cs.CV · cs.GR· cs.LG

Discretizing Group-Convolutional Neural Networks for 3D Geometry in Feature Space

Daniel Franzen , Jean Philip Filling , Michael Wand This is my paper

Pith reviewed 2026-05-19 15:52 UTC · model grok-4.3

classification 💻 cs.CV cs.GRcs.LG
keywords group-convolutional neural networks3D geometryfeature space samplingequivariancediscretizationclassification accuracy3D classifiers
0
0 comments X p. Extension
pith:KXHWQWJR Add to your LaTeX paper What is a Pith Number? 
\usepackage{pith}
\pithnumber{KXHWQWJR}

Prints a linked pith:KXHWQWJR badge after your title and writes the identifier into PDF metadata. Compiles on arXiv with no extra files. Learn more

The pith

Sampling transformations in feature space preserves accuracy in group-convolutional networks for 3D geometry while reducing costs.

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

Group-convolutional neural networks maintain equivariance by densely sampling a transformation group such as rotations and translations and correlating data with filters in each pose. The paper replaces this geometric density with representative samples chosen by feature similarity, which decouples resolution from memory and compute demands. Empirical results show that even coarse feature-space sampling keeps classification accuracy high on 3D tasks, enabling precomputation of similar poses and substantially faster training of equivariant classifiers.

Core claim

Replacing geometrically dense samples with representative samples selected by feature similarity decouples geometric resolution from memory and processing costs during training and inference. This provides a novel way to trade off computational effort and accuracy, with the main empirical result that coarse feature-space sampling already preserves classification accuracy remarkably well and permits precomputation based on geometric similarity.

What carries the argument

Feature-space sampling that replaces dense sampling of the transformation group with representative samples chosen according to feature similarity.

If this is right

  • Training of equivariant 3D classifiers becomes substantially faster through precomputation of geometrically similar poses.
  • Memory and processing costs are decoupled from the number of degrees of freedom in the transformation group.
  • Equivariance can be maintained at lower computational expense by trading geometric density for feature-based selection.
  • Practical deployment of GCNNs on 3D data becomes more feasible as exponential cost growth with group size is avoided.

Where Pith is reading between the lines

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

  • The same feature-space approach could extend to other transformation groups or higher-dimensional data where dense sampling is prohibitive.
  • Precomputed similarity tables might allow dynamic adjustment of sampling density during inference based on available compute.
  • Limits of the method could be tested by measuring equivariance error directly rather than downstream classification accuracy alone.

Load-bearing premise

Representative samples selected purely by feature similarity are sufficient to maintain the equivariance properties and accuracy that dense geometric sampling would have provided.

What would settle it

A significant drop in classification accuracy on standard 3D benchmarks when replacing dense geometric sampling with the proposed coarse feature-space sampling would falsify the central claim.

Figures

Figures reproduced from arXiv: 2605.15368 by Daniel Franzen, Jean Philip Filling, Michael Wand.

Figure 1
Figure 1. Figure 1: Typical 3D shapes (a) have a lot of locally (approximately) similar geometry (b). We detect similar regions (c,d) and compress the computations [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation of our translation-only equivariant classification architecture on the ModelNet40 test set. (a) shows the accuracy for different augmentation [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Evaluation of our translation-only equivariant classification architecture on the ScanObjectNN test set. Same plots as in Fig. 2. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Evaluation of our surface-normal-aligned [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation of our surface-normal-aligned [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of clustering strategies on ScanObjectNN (16384 points, 64 representatives, translation-only equivariant model). The ablation study shows [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results for our segmentation experiment on the ScanNet20 validation set (not using color information). Note that the plots report mean class IoU [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
read the original abstract

Group-convolutional neural networks (GCNNs) are among the most important methods for introducing symmetry as an inductive bias in deep learning: In each linear layer, GCNNs sample a transformation group $G$ densely and correlate data and filters in different poses (with suitable anti-aliasing for steerable GCNNs) to maintain equivariance with respect to $G$. Unfortunately, applying filters to many data items resulting from this sampling is expensive (even for translations alone, i.e., in ordinary CNNs), and costs grow exponentially with increasing degrees of freedom (such as translations and rotations in 3D), which often hinders practical applications. In this paper, we propose sampling in feature space, i.e., replacing geometrically dense samples with representative samples selected by feature similarity. This decouples geometric resolution from memory and processing costs during training and inference, providing a novel way to trade off computational effort and accuracy. Our main empirical finding is that a coarse feature-space sampling already preserves classification accuracy remarkably well, which permits precomputation based on geometric similarity, accelerating the training of equivariant 3D classifiers substantially.

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 replacing dense geometric sampling in group-convolutional neural networks (GCNNs) for 3D geometry with representative samples selected by feature similarity in feature space. This decouples geometric resolution from computational costs during training and inference. The central empirical claim is that coarse feature-space sampling already preserves classification accuracy remarkably well, enabling precomputation based on geometric similarity and substantial acceleration of equivariant 3D classifiers.

Significance. If the empirical result holds and the discretization maintains sufficient equivariance, the approach would offer a practical efficiency gain for symmetry-aware networks in 3D, where dense sampling costs grow rapidly with degrees of freedom. It introduces a tunable trade-off between accuracy and compute that could broaden adoption of GCNNs in resource-constrained settings.

major comments (2)
  1. Abstract: the central claim that coarse feature-space sampling preserves classification accuracy 'remarkably well' is stated without any reference to datasets, baselines, error bars, or the precise definition of feature similarity. This absence prevents verification of the claim's support and generalizability from the provided text.
  2. The method replaces geometrically dense sampling (with anti-aliasing) by feature-similarity selection, which decouples the discretization from the group structure G. For the accuracy-preservation claim to be load-bearing rather than task-specific, the manuscript must demonstrate that the selected points still induce a representation equivariant (or sufficiently close) under the original group actions; otherwise the observed accuracy may not generalize beyond the particular datasets.
minor comments (1)
  1. The abstract would be strengthened by a single sentence indicating the 3D classification tasks or data modalities on which the empirical result was observed.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful review and constructive suggestions. We address each of the major comments below and outline the revisions we will make to the manuscript.

read point-by-point responses

  1. Referee: Abstract: the central claim that coarse feature-space sampling preserves classification accuracy 'remarkably well' is stated without any reference to datasets, baselines, error bars, or the precise definition of feature similarity. This absence prevents verification of the claim's support and generalizability from the provided text.

    Authors: We agree that the abstract would benefit from greater specificity to allow readers to better assess the claim. In the revised version, we will update the abstract to reference the datasets used (such as ModelNet for 3D point cloud classification), the baselines compared against (dense geometric sampling), and note that feature similarity is defined using distances in the embedding space of a backbone network. We will also mention that results include standard deviations from repeated experiments where applicable. revision: yes

  2. Referee: The method replaces geometrically dense sampling (with anti-aliasing) by feature-similarity selection, which decouples the discretization from the group structure G. For the accuracy-preservation claim to be load-bearing rather than task-specific, the manuscript must demonstrate that the selected points still induce a representation equivariant (or sufficiently close) under the original group actions; otherwise the observed accuracy may not generalize beyond the particular datasets.

    Authors: This comment raises a crucial point about the preservation of equivariance. Our approach selects samples based on feature similarity to approximate the dense sampling while reducing computational cost. Although the manuscript focuses on empirical performance, we acknowledge the need for more explicit analysis. We will add a new section or subsection that quantifies the equivariance error by measuring the difference in network outputs under group transformations before and after sampling. This will help demonstrate that the approximation remains sufficiently equivariant for the tasks considered. We believe this addition will address concerns about generalizability. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical discretization trade-off is self-contained

full rationale

The paper's central contribution is an empirical proposal to replace dense geometric sampling in GCNNs with representative samples chosen by feature similarity, with the key observation that coarse feature-space sampling preserves classification accuracy on the tested tasks. No equations, derivations, or predictions are presented that reduce by construction to fitted inputs, self-definitions, or self-citation chains; the method is explicitly framed as a practical accuracy-compute trade-off rather than a closed-form identity or uniqueness result. The full manuscript contains no load-bearing steps where an output is forced to match its inputs via the paper's own definitions or prior author work, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; assessment is limited to the high-level proposal.

pith-pipeline@v0.9.0 · 5734 in / 995 out tokens · 37334 ms · 2026-05-19T15:52:59.929369+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

68 extracted references · 68 canonical work pages · 3 internal anchors

  1. [1]

    Matan Atzmon, Haggai Maron, and Yaron Lipman. 2018. Point Convolutional Neural Networks by Extension Operators. Issue 4

    work page 2018

  2. [2]

    Erik J. Bekkers. 2020. B-Spline CNNs on Lie Groups. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=H1gBhkBFDH

    work page 2020

  3. [3]

    Bekkers, Sharvaree Vadgama, Rob D

    Erik J. Bekkers, Sharvaree Vadgama, Rob D. Hesselink, Putri A. van der Linden, and David W. Romero. 2024. Fast, Expressive SE(𝑛) Equivariant Networks through Weight-Sharing in Position-Orientation Space. InInternational Conference on Learning Representations (ICLR). https://openreview.net/forum?id=dPHLbUqGbr

    work page 2024

  4. [4]

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Fe- ichtenhofer, and Judy Hoffman. 2023. Token Merging: Your ViT But Faster. InThe Eleventh International Conference on Learning Representations. https: //openreview.net/forum?id=JroZRaRw7Eu

    work page 2023

  5. [5]

    Johann Brehmer, Sönke Behrends, Pim de Haan, and Taco Cohen. 2024. Does Equivariance Matter at Scale?arXiv preprint arXiv:2410.23179(2024). arXiv:2410.23179 [cs.LG]

    work page arXiv 2024

  6. [6]

    Johann Brehmer, Pim De Haan, Sönke Behrends, and Taco Cohen. 2023. Geo- metric Algebra Transformer. InThirty-seventh Conference on Neural Information Processing Systems. https://openreview.net/forum?id=M7r2CO4tJC

    work page 2023

  7. [7]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. 2021. Geo- metric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges.arXiv preprint arXiv:2104.13478(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  8. [8]

    Qendrim Bytyqi, Nicola Wolpert, and Elmar Schömer. 2020. Local-Area- Learning Network: Meaningful Local Areas for Efficient Point Cloud Analysis. arXiv:2006.07226 [cs.CV] https://arxiv.org/abs/2006.07226

    work page arXiv 2020

  9. [9]

    Gabriele Cesa, Leon Lang, and Maurice Weiler. 2022. A Program to Build E(N)- Equivariant Steerable CNNs. InInternational Conference on Learning Representa- tions. https://openreview.net/forum?id=WE4qe9xlnQw

    work page 2022

  10. [10]

    Haiwei Chen, Shichen Liu, Weikai Chen, Hao Li, and Randall Hill. 2021. Equi- variant Point Network for 3D Point Cloud Analysis. (2021), 14514–14523

    work page 2021

  11. [11]

    Ta-Ying Cheng, Qingyong Hu, Qian Xie, Niki Trigoni, and Andrew Markham

    work page

  12. [12]

    InEuropean Conference on Computer Vision (ECCV)

    Meta-Sampler: Almost Universal yet Task-Oriented Sampling for Point Clouds. InEuropean Conference on Computer Vision (ECCV)

    work page

  13. [13]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. 2016. Group Equivariant Convolutional Net- works. InProceedings of the International Conference on Machine Learning (ICML)

    work page 2016

  14. [14]

    Cohen and Max Welling

    Taco S. Cohen and Max Welling. 2017. Steerable CNNs. InInternational Conference on Learning Representations (ICLR)

    work page 2017

  15. [15]

    Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. 2017. ScanNet: Richly-Annotated 3D Reconstructions of Indoor Scenes. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 5828–5839

    work page 2017

  16. [16]

    Pim de Haan, Maurice Weiler, Taco Cohen, and Max Welling. 2021. Gauge Equivariant Mesh CNNs: Anisotropic Convolutions on Geometric Graphs. In International Conference on Learning Representations (ICLR). https://openreview. net/forum?id=Jnspzp-oIZE

    work page 2021

  17. [17]

    DeepSeek-AI. 2025. DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention. https://github.com/deepseek-ai/DeepSeek-V3.2- Exp/blob/main/DeepSeek_V3_2.pdf

    work page 2025

  18. [18]

    Fei and Z

    J. Fei and Z. Deng. 2024. Rotation invariance and equivariance in 3D deep learning: a survey.Artif Intell Rev57, 168 (2024). doi:10.1007/s10462-024-10741-2

    work page doi:10.1007/s10462-024-10741-2 2024

  19. [19]

    Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Müller. 2018. SplineCNN: Fast Geometric Deep Learning with Continuous B-Spline Kernels. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2018

  20. [20]

    Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Ragu- ram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazeb- nik, and Marc Pollefeys. 2010. Building Rome on a Cloudless Day. InComputer Vision – ECCV 2010, Kostas Daniilidis, Petros Maragos, and Nikos Paragios (Eds.). Springer Berlin Heidelberg, Berlin, Heidelbe...

    work page 2010

  21. [21]

    Fuchs, Daniel E

    Fabian B. Fuchs, Daniel E. Worrall, Volker Fischer, and Max Welling. 2020. SE(3)- transformers: 3D roto-translation equivariant attention networks. InProceedings of the 34th International Conference on Neural Information Processing Systems (Vancouver, BC, Canada)(NIPS ’20). Curran Associates Inc., Red Hook, NY, USA, Article 166, 12 pages

    work page 2020

  22. [22]

    Natasha Gelfand and Leonidas J. Guibas. 2004. Shape segmentation using local slippage analysis. InProceedings of the 2004 Eurographics/ACM SIGGRAPH Sympo- sium on Geometry Processing(Nice, France)(SGP ’04). Association for Computing Machinery, New York, NY, USA, 214–223. doi:10.1145/1057432.1057461

    work page doi:10.1145/1057432.1057461 2004

  23. [23]

    2016.Deep Learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. 2016.Deep Learning. MIT Press. http://www.deeplearningbook.org

    work page 2016

  24. [24]

    Martin, and Shi-Min Hu

    Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R. Martin, and Shi-Min Hu. 2021. PCT: Point cloud transformer.Computational Visual Media 7, 2 (Apr 2021), 187–199. doi:10.1007/s41095-021-0229-5

    work page doi:10.1007/s41095-021-0229-5 2021

  25. [25]

    Rana Hanocka, Amir Hertz, N. Fish, R. Giryes, S. Fleishman, and D. Cohen-Or

    work page

  26. [26]

    MeshCNN: a network with an edge.ACM Transactions on Graphics (TOG) 38 (2019), 1 – 12

    work page 2019

  27. [27]

    Sergey Ioffe and Christian Szegedy. 2015. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. InProceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015 (JMLR Workshop and Conference Proceedings, Vol. 37), Francis R. Bach and David M. Blei (Eds.). JMLR.org, 448–456

    work page 2015

  28. [28]

    Wessels, Friso de Kruiff, Thijs P

    Mohammad Mohaiminul Islam, Rishabh Anand, David R. Wessels, Friso de Kruiff, Thijs P. Kuipers, Rex Ying, Clara I. Sánchez, Sharvaree Vadgama, Georg Bökman, and Erik J. Bekkers. 2025. Platonic Transformers: A Solid Choice For Equivariance. arXiv preprint arXiv:2510.03511(2025). arXiv:2510.03511 [cs.CV]

    work page arXiv 2025

  29. [30]

    Kerber, M

    J. Kerber, M. Bokeloh, M. Wand, and H.-P. Seidel. 2013. Scalable Sym- metry Detection for Urban Scenes.Computer Graphics Forum32, 1 (2013), 3–15. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1467- 8659.2012.03226.x doi:10.1111/j.1467-8659.2012.03226.x

    work page doi:10.1111/j.1467- 2013

  30. [31]

    Knigge, David W

    David M. Knigge, David W. Romero, and Erik J. Bekkers. 2022. Exploiting Redun- dancy: Separable Group Convolutional Networks on Lie Groups. InProceedings of the 39th International Conference on Machine Learning (ICML) (Proceedings of Machine Learning Research, Vol. 162). PMLR, 11359–11386

    work page 2022

  31. [32]

    Risi Kondor, Zhen Lin, and Shubhendu Trivedi. 2018. Clebsch–Gordan nets: a fully fourier space spherical convolutional neural network. InProceedings of the 32nd International Conference on Neural Information Processing Systems(Montréal, Canada)(NIPS’18). Curran Associates Inc., Red Hook, NY, USA, 10138–10147

    work page 2018

  32. [33]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. ImageNet Clas- sification with Deep Convolutional Neural Networks. InAdvances in Neural Information Processing Systems, F. Pereira, C.J. Burges, L. Bottou, and K.Q. Wein- berger (Eds.), Vol. 25. Curran Associates, Inc. https://proceedings.neurips.cc/ paper_files/paper/2012/file/c399862d3b9d6b76...

    work page 2012

  33. [34]

    Kuipers and Erik J

    Thijs P. Kuipers and Erik J. Bekkers. 2023. Regular SE(3) Group Convolutions for Volumetric Medical Image Analysis. InMedical Image Computing and Computer Assisted Intervention (MICCAI) (Lecture Notes in Computer Science, Vol. 14222). Springer, 252–261

    work page 2023

  34. [35]

    Taeyoung Lee. 2022. Real Harmonic Analysis on the Special Orthogonal Group. Int. J. Anal. Appl.20, 21 (2022)

    work page 2022

  35. [36]

    Yangyan Li, Rui Bu, M. Sun, W. Wu, X. Di, and B. Chen. 2018. PointCNN: Convo- lution On X-Transformed Points.arXiv: Computer Vision and Pattern Recognition (2018)

    work page 2018

  36. [37]

    Zhuangzi Li, Shan Liu, Wei Gao, Guanbin Li, and Ge Li. 2025. S4R: Rethink- ing Point Cloud Sampling via Guiding Upsampling-Aware Perception.IEEE Transactions on Multimedia27 (2025), 6677–6689. doi:10.1109/TMM.2025.3586148

    work page doi:10.1109/tmm.2025.3586148 2025

  37. [38]

    Jiehong Lin, Hongyang Li, Ke Chen, Jiangbo Lu, and Kui Jia. 2021. Sparse Steerable Convolutions: An Efficient Learning of SE(3)-Equivariant Features for Estimation and Tracking of Object Poses in 3D Space. InAdvances in Neural Information Processing Systems, A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan (Eds.). https://openreview.net/forum?...

    work page 2021

  38. [39]

    Mitra, Leonidas J

    Niloy J. Mitra, Leonidas J. Guibas, and Mark Pauly. 2006. Partial and approximate symmetry detection for 3D geometry.ACM Trans. Graph.25, 3 (July 2006), 560–568. doi:10.1145/1141911.1141924

    work page doi:10.1145/1141911.1141924 2006

  39. [40]

    Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan

    Niloy J. Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. 2013. Symmetry in 3D Geometry: Extraction and Applications.Computer Graphics Forum32, 6 (2013), 1–23. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.12010 doi:10.1111/cgf.12010

    work page doi:10.1111/cgf.12010 2013

  40. [41]

    Ehsan Nezhadarya, Ehsan Taghavi, Ryan Razani, Bingbing Liu, and Jun Luo

    work page

  41. [42]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Adaptive Hierarchical Down-Sampling for Point Cloud Classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page

  42. [43]

    Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. 2022. Masked autoencoders for point cloud self-supervised learning. In Computer Vision–ECCV 2022: 17th European Conference, Tel A viv, Israel, October Discretizing Group-Convolutional Neural Networks for 3D Geometry in Feature Space•11 23–27, 2022, Proceedings, Part II. Spring...

    work page 2022

  43. [44]

    Qi, Hao Su, Kaichun Mo, and L

    C. Qi, Hao Su, Kaichun Mo, and L. Guibas. 2017. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)(2017), 77–85

    work page 2017

  44. [45]

    C. Qi, L. Yi, Hao Su, and L. Guibas. 2017. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. InNIPS

    work page 2017

  45. [46]

    Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J

    Charles R. Qi, Hao Su, Matthias Nießner, Angela Dai, Mengyuan Yan, and Leonidas J. Guibas. 2016. Volumetric and Multi-view CNNs for Object Clas- sification on 3D Data. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 5648–5656. doi:10.1109/CVPR.2016.609

    work page doi:10.1109/cvpr.2016.609 2016

  46. [47]

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh

    work page

  47. [48]

    InAdvances in Neural Information Processing Systems, A

    DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsi- fication. InAdvances in Neural Information Processing Systems, A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan (Eds.). https://openreview.net/ forum?id=kR95DuwwXHZ

    work page

  48. [49]

    David Ruhe, Johannes Brandstetter, and Patrick Forré. 2023. Clifford Group Equivariant Neural Networks. InThirty-seventh Conference on Neural Information Processing Systems. https://openreview.net/forum?id=n84bzMrGUD

    work page 2023

  49. [50]

    Víctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. 2021. E(n) Equivari- ant Graph Neural Networks. InProceedings of the 38th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 139), Marina Meila and Tong Zhang (Eds.). PMLR, 9323–9332. https://proceedings.mlr.press/ v139/satorras21a.html

    work page 2021

  50. [51]

    Tkatchenko, and K

    Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, A. Tkatchenko, and K. Müller. 2017. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. InNIPS

    work page 2017

  51. [52]

    Shuran Song and Jianxiong Xiao. 2014. Sliding Shapes for 3D Object Detection in Depth Images. InComputer Vision – ECCV 2014, David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars (Eds.). Springer International Publishing, Cham, 634–651

    work page 2014

  52. [53]

    Vishal Thengane, Xiatian Zhu, Salim Bouzerdoum, Son Lam Phung, and Yunpeng Li. 2025. Foundational models for 3d point clouds: A survey and outlook.arXiv preprint arXiv:2501.18594(2025)

    work page arXiv 2025

  53. [54]

    Thomas, C

    H. Thomas, C. Qi, Jean-Emmanuel Deschaud, B. Marcotegui, F. Goulette, and L. Guibas. 2019. KPConv: Flexible and Deformable Convolution for Point Clouds. 2019 IEEE/CVF International Conference on Computer Vision (ICCV)(2019), 6410– 6419

    work page 2019

  54. [55]

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

    Nathaniel Thomas, Tess Smidt, Steven M. Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. 2018. Tensor Field Networks: Rotation- and Translation-Equivariant Neural Networks for 3D Point Clouds.CoRR abs/1802.08219 (2018). arXiv:1802.08219 http://arxiv.org/abs/1802.08219

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  55. [56]

    Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. 2019. Revisiting Point Cloud Classification: A New Benchmark Dataset and Classification Model on Real-World Data. InInternational Conference on Computer Vision (ICCV)

    work page 2019

  56. [57]

    Sharvaree Vadgama, Mohammad Mohaiminul Islam, Domas Buracas, Christian Shewmake, Artem Moskalev, and Erik J. Bekkers. 2025. Probing Equivariance and Symmetry Breaking in Convolutional Networks. InAdvances in Neural Infor- mation Processing Systems (NeurIPS). arXiv:2501.01999 [cs.CV]

    work page arXiv 2025

  57. [58]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems, I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30. Curran Associates, Inc. https://proc...

    work page 2017

  58. [59]

    2010.The nonequispaced fast SO(3) Fourier transform, generali- sations and applications

    Antje Vollrath. 2010.The nonequispaced fast SO(3) Fourier transform, generali- sations and applications. Ph. D. Dissertation. Universität zu Lübeck, Institut für Mathematik

    work page 2010

  59. [60]

    Jiahua Wang, Yao Zhao, Ting Liu, and Shikui Wei. 2021. GDS: Global Description Guided Down-Sampling for 3D Point Cloud Classification. InProceedings of the 2020 4th International Conference on Vision, Image and Signal Processing(Bangkok, Thailand)(ICVISP 2020). Association for Computing Machinery, New York, NY, USA, Article 19, 6 pages. doi:10.1145/344882...

    work page doi:10.1145/3448823.3448842 2021

  60. [61]

    Sarma, Michael M

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. 2019. Dynamic Graph CNN for Learning on Point Clouds. ACM Transactions on Graphics (TOG)(2019)

    work page 2019

  61. [62]

    Maurice Weiler, Wouter Boomsma, Mario Geiger, Max Welling, and Taco S. Co- hen. 2018. 3D Steerable CNNs, Learning Rotationally Equivariant Features in Volumetric Data. InAdvances in Neural Information Processing Systems (NeurIPS)

    work page 2018

  62. [63]

    Maurice Weiler, Patrick Forré, Erik Verlinde, and Max Welling. 2021. Coordi- nate Independent Convolutional Networks — Isometry and Gauge Equivariant Convolutions on Riemannian Manifolds.arXiv preprint arXiv:2106.06020(2021). arXiv:2106.06020 [cs.LG]

    work page arXiv 2021

  63. [64]

    Ruben Wiersma, Elmar Eisemann, and Klaus Hildebrandt. 2020. CNNs on Surfaces using Rotation-Equivariant Features.Transactions on Graphics39, 4 (July 2020). doi:10.1145/3386569.3392437

    work page doi:10.1145/3386569.3392437 2020

  64. [65]

    Marysia Winkels and Taco S. Cohen. 2018. 3D G-CNNs for Pulmonary Nodule Detection.arXiv preprint arXiv:1804.04656(2018). arXiv:1804.04656 [cs.CV]

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  65. [66]

    Daniel Worrall and Gabriel Brostow. 2018. CubeNet: Equivariance to 3D Rotation and Translation. InProceedings of the European Conference on Computer Vision (ECCV). 567–584

    work page 2018

  66. [67]

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 2015. 3d shapenets: A deep representation for volumet- ric shapes. InProceedings of the IEEE conference on computer vision and pattern recognition. 1912–1920

    work page 2015

  67. [68]

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. 2021. Point Transformer. In2021 IEEE/CVF International Conference on Computer Vision (ICCV). 16239–16248. doi:10.1109/ICCV48922.2021.01595

    work page doi:10.1109/iccv48922.2021.01595 2021

  68. [69]

    Minghan Zhu, Maani Ghaffari, William A Clark, and Huei Peng. 2023. E2PN: Effi- cient SE (3)-equivariant point network. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 1223–1232

    work page 2023