Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Hyperbolic Contrastive Learning for Hierarchical 3D Point Cloud Embedding

T0 review · 5 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A training strategy extends hyperbolic contrastive pretraining to 3D point clouds, using entailment, centroid, and alignment regularizers to enforce a text-image-point-cloud hierarchy, and reports improved classification and segmentation.

desk verdict A sensible extension of hyperbolic image-text pretraining to 3D point clouds, but the central regularizer equations are unusable as printed. read the letter →

arxiv 2501.02285 v2 pith:RGQ5RTUU submitted 2025-01-04 cs.CV cs.AI

Yingjie Liu , Pengyu Zhang , Ziyao He , Mingsong Chen , Xuan Tang , Xian Wei This is my paper · ORCID
classification cs.CVcs.AI
keywords hyperbolicgeometrypointcloudcontrastivelearninghierarchicalembeddingsentailmentlossLorentzhyperboloid3Drepresentationmulti-modalpretraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that 3D point cloud embeddings can be trained inside a hyperbolic space together with text and image embeddings, and that making the hierarchy explicit improves downstream 3D tasks. It extends a reconstruction-guided contrastive point cloud encoder with three regularizers: an entailment loss that forces images into the cones of point clouds and point clouds into the cones of text, a centroid loss that orders the modality centroids by abstraction level, and an alignment loss that captures whole-to-part relations within point clouds. The paper reports that the resulting embeddings are more tree-like by a $\delta$-hyperbolicity measure and that they beat CLIP-based, MERU-based, and ReCon baselines on ModelNet40 classification and ShapeNetPart part segmentation. A sympathetic reader would care because the question is whether geometric priors that help language-image models transfer to a third, geometrically richer modality.

What carries the argument

The load-bearing object is the hyperbolic entailment regularizer, adapted from the MERU formulation: text, image, and point cloud embeddings live on a Lorentz hyperboloid, and the loss penalizes cases where the sum of exterior angles between image-point cloud and point cloud-text pairs exceeds the cone half-apertures of the parents. The half-aperture $\operatorname{aper}(x)$ and exterior angle $\operatorname{ext}(x,y)$ are the formulas that carry this mechanism, with an Einstein-midpoint centroid term $\mathcal{L}_{\mathrm{cent}}$ ordering modality centroids and an alignment term that runs the point cloud encoder twice, on full and masked inputs, to supervise a part embedding. The three regularizers are combined with the reconstruction and contrastive losses using uncertainty-based weighting.

What would settle it

Numerically evaluate the half-aperture formula on two points of the Lorentz hyperboloid that lie on the same ray with different distances from the origin: as printed, the formula yields a larger half-aperture for the larger distance, which contradicts the intended cone geometry; re-running the experiments with a corrected formula would settle whether the reported hierarchy gains persist.

Watch

Extended reading notes

Core claim

The central claim is that adding hyperbolic hierarchy regularizers to a reconstruction-guided contrastive pretraining produces point cloud embeddings with an explicit partial order across text, images, and 3D point clouds, and that this order is measurable and useful. On the paper's terms, the learned embeddings have lower $\delta$-hyperbolicity than the teacher targets, point cloud centroids sit in the intended order relative to text and image centroids, and the whole-to-part structure inside point clouds is visible in distance distributions. The paper's experiments report the modified approach reaching 93.64 percent accuracy on ModelNet40 with 1k points, 95.37 percent on ModelNet10, and 84.91 mean class IoU on ShapeNetPart, all above the compared baselines.

Load-bearing premise

The load-bearing premise is that the paper's formula for the cone half-angle, which depends on a distance from the origin and an undefined constant $K$, correctly describes the loss that was actually trained; as printed, that formula gives larger cones for points farther from the origin, the opposite of the intended hierarchy, so if the printed formula is wrong the claimed hierarchical ordering is not supported by the equations.

Editorial extensions

If this is right

  • Fine-tuned classification on ModelNet40 improves to 93.64 percent at 1k points, above the CLIP-based 93.44 and MERU-based 93.07 baselines, without test-time voting.
  • Part segmentation on ShapeNetPart reaches 84.91 mean class IoU, the best average in the paper's comparison table.
  • The measured hyperbolicity of text embeddings drops to 0.1716 with the modified regularizers versus 0.2695 for the CLIP-based variant, indicating more tree-like structure.
  • The centroid regularizer keeps a detectable modality gap while ordering point cloud centroids between text and image centroids, which the paper links to preserved hierarchy.
  • Sparse dictionary analysis shows roughly 30 percent of atoms frequently activated with class-distinct patterns, suggesting the hierarchical embeddings are also more disentangled.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the printed aperture formula is taken literally, $\operatorname{aper}(x)$ grows with the embedding's distance from the origin, which would make specific concepts have larger cones than abstract ones, the reverse of the intended geometry; the reported gains may therefore be carried mainly by the centroid and alignment losses rather than the entailment term.
  • This can be tested directly: compute $\operatorname{aper}(x)$ for two points on the same ray of the hyperboloid at different radii; the formula as printed gives a larger half-aperture for the larger radius, and correcting the formula (or defining the constant $K$) would change what loss was actually trained.
  • Replacing the frozen hyperbolic teacher with a Euclidean teacher while keeping the same regularizers would isolate how much of the hierarchy gain comes from the hyperbolic geometry as opposed to the extra losses.
  • The dictionary-learning evidence suggests the embeddings may decompose into interpretable part-like atoms, but the paper does not yet connect that decomposition to the classification or segmentation gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes extending hyperbolic multi-modal contrastive pre-training to 3D point clouds. The method builds on the reconstruction-guided contrastive framework of ReCon and a frozen MERU teacher, and adds three families of regularizers: a hyperbolic entailment regularizer intended to impose a text->image->point-cloud cone hierarchy, a centroid regularizer intended to enforce inter-modal distance ordering, and an alignment regularizer intended to capture whole-to-part structure within the point-cloud modality. The authors report experiments on ShapeNetPart part segmentation, ModelNet40/ModelNet10 fine-tuned classification, and ModelNet40 few-shot classification, together with hyperbolicity and embedding-distance analyses aimed at demonstrating hierarchical structure. The central novelty is the set of hyperbolic regularizers, and the paper claims that these regularizers improve downstream point-cloud tasks.

Significance. If the proposed regularizers are well-defined and actually produce hierarchical hyperbolic point-cloud embeddings, the work would be a useful step toward bringing 3D point clouds into hyperbolic vision-language pre-training. The paper addresses a genuine gap, since most existing hyperbolic contrastive work covers text and images only. It also correctly identifies the value of downstream external benchmarks, which keep the main claim partially independent of the hierarchy analyses. However, the significance is currently limited by reproducibility problems in the core loss definitions and by the small, unquantified downstream gains. The strongest parts are the experimental setup on standard benchmarks and the use of Lorentz-geometric tools; the weakest parts are the garbled equations for the two new regularizers and the missing alignment-loss term.

major comments (5)
  1. [Section IV-B, aperture formula] The half-aperture is printed as aper(x) = sin^{-1}(2K\sqrt{c} ||xspace||) with the constraint ||xspace|| >= 2K\sqrt{c}. As written, this grows with ||xspace||, whereas the intended entailment-cone geometry requires narrower apertures for embeddings farther from the origin; the constant K is never defined. If this is a typesetting error and the intended expression is, e.g., sin^{-1}(2K/(\sqrt{c} ||xspace||)), the equation and the accompanying constraint must be corrected and K specified before the loss can be evaluated or reproduced.
  2. [Section IV-B, centroid regularizer] After defining p, q, r by equations such as p = d_L(O, p) = ..., the loss is written as L_cent = ||Ze - p||^2 + ||Ye - q||^2 + ||Xe - r||^2. As written, this subtracts scalar distances from embedding vectors, so the term is not evaluable. The intended objective must be rewritten with distinct notation for the target scalar distances and the embedding centroids, and the correspondence between p, q, r and the text/point-cloud/image centroids should be stated explicitly.
  3. [Section IV-C and Section IV-D] Section IV-C is titled 'Alignment Loss and Hyperbolicity Analysis' and describes running the point-cloud encoder on full and masked point clouds, but it never provides an equation for an alignment loss. The total loss in Section IV-D is L = sum_i (e^{-s_i} L_i + s_i) with L_i in {L_cent, L_entail, L_Rec, L_Con}, and no alignment term appears. Either the alignment loss must be written out and included in the total loss, or the claim that an alignment regularizer is part of the method should be withdrawn.
  4. [Section V-A, Figure 2] The hierarchy evidence in Figure 2 is partly circular: the model is trained with L_entail and L_cent, which explicitly enforce entailment between text, image, and point-cloud cones and enforce centroid distance ordering. Observing those same orderings afterward is therefore a restatement of the training objective. The demonstration would be meaningful if the authors also report the distance distributions for the unregularized CLIP-based and MERU-based baselines and include an ablation that removes only the hierarchy regularizers, showing what structure emerges without explicit enforcement.
  5. [Tables II, III, and IV] The downstream improvements over the non-hierarchical baselines are small, e.g., ShapeNetPart mIoUC 84.91 vs 84.70 for the MERU-based baseline and ModelNet40 1k accuracy 93.64 vs 93.44 for CLIP-based, and no error bars, multiple seeds, or significance tests are reported. Given that the paper's central claim is that the hierarchical embeddings significantly improve downstream tasks, please provide variance estimates and, ideally, ablations separating L_entail and L_cent from the ReCon baseline so the contribution of each regularizer can be assessed.
minor comments (5)
  1. [Table I] The MERU modified-based row lists only the Text value (0.1716 +/- 0.02735); the Image and Point Clouds entries appear to be missing, so the table is incomplete as presented.
  2. [References and Section IV-A] The paper sometimes calls MERU [6] and sometimes cites [37] for MERU, e.g., Section IV-A says 'pre-trained hyperbolic language-image model MERU [37]', but [37] is the modality-gap paper, not MERU. Please disambiguate and cite consistently.
  3. [Section IV-B, centroid regularizer paragraph] There is a duplicated article in 'and the the centroid takes the following form'; please correct this typo.
  4. [Section V-A, figure references] The text for the whole-to-part distance analysis refers to Figure 3a, while the corresponding panel appears to be Figure 2a; please align the figure citations with the captions.
  5. [Section IV-B, notation] The norm in L_cent is written as 'where is the Euclidean norm'; the symbol for the norm is missing and should be supplied.

Circularity Check

1 steps flagged · score 4.0 of 10

Hierarchy evidence in RQ1 is a direct read-out of the training losses; downstream benchmarks remain external.

  1. fitted input called prediction [Section IV-B (L_cent definition) and Section V-A (Hierarchical Embedding Analysis, Figure 2b)]
    "p > q > r >0 to ensure that the centroid relationships ... we ensure that the centroid of text embeddings is closer to the origin than the centroid of visual embeddings, and the centroid of 3D Point Cloud embeddings should be closer to the origin than the centroid of the 2D image embeddings ... Figure 2b shows that our regularizers, named our MERU (modified)-based method explicitly construct hierarchical relations across intra-modal embeddings while maintaining their modality gap"

    The training objective L explicitly includes L_cent (which imposes p > q > r on the centroid distances) and L_entail (which imposes cone containment among text, image, and point-cloud embeddings). The RQ1 analysis then plots distances from the origin and reports that the regularizers 'explicitly construct hierarchical relations.' Demonstrating that a model possesses a property its loss was designed to force is a sanity check, not an independent prediction: the observed hierarchy is equivalent to the optimization constraint by construction. The downstream ModelNet/ShapeNetPart evaluations are external and non-circular, so the circularity is partial and confined to the hierarchy demonstration.

full rationale

The paper's central downstream claims (fine-tuned classification, few-shot learning, part segmentation) are evaluated on external benchmarks and do not reduce to the training losses, so those results are independent evidence. The circular component is limited to RQ1, where the paper presents the centroid ordering and cone containment produced by its own regularizers as evidence of hierarchical embedding quality; this is the training objective read back out of the model. The self-citation to [20] (a baseline by overlapping authors) is not load-bearing. The undefined aperture constant K, the inverted aperture formula, and the scalar-vector mismatch in L_cent are correctness and reproducibility concerns, not circularity, so they do not further raise the circularity score. Overall, partial circularity in the hierarchy analysis, but the main performance claims stand on external evaluations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several modeling choices that are imposed by the regularizers rather than derived or measured. The aperture equation itself is suspect, and the hierarchy evidence partly reflects the training objective. No new physical or conceptual entities are introduced.

free parameters (3)
  • alpha_pts (point cloud embedding scale) = learned, value not reported
    This scale controls the norm of point cloud embeddings before the exponential map and is learned during pre-training, as described in Implementation Details.
  • learnable loss weights s_i = learned, values not reported
    The homoscedastic uncertainty weights for L_cent, L_entail, L_Rec, and L_Con are learned and shape the joint objective; no final values or sensitivity analysis are provided.
  • K in aperture formula = not specified
    The constant K appears in the aperture formula of Section IV-B but is never defined. If it is inherited from MERU, it is not stated; as written it is an unexplained free constant.
assumptions (4)
  • ad hoc to paper Entailment cones in Lorentz space represent semantic hierarchy, with images entailed within point-cloud cones and point clouds entailed within text cones.
    Section IV-B constructs this partial order as a design choice; no independent evidence is given that this ordering holds for all text-image-point cloud pairs.
  • ad hoc to paper The centroid distance ordering p > q > r is the correct cross-modal hierarchy.
    The centroid loss L_cent is designed to enforce this ordering, making it an imposed assumption rather than a measured fact.
  • ad hoc to paper Running the encoder on full and masked point clouds captures whole-to-part hierarchy.
    The alignment regularizer relies on this assumption, which is plausible but not independently validated.
  • domain assumption Gromov delta-hyperbolicity computed on 128-sample batches reflects the global embedding space.
    Section V-A computes delta-rel on individual batches and averages over 409 batches; the paper assumes this subsampling represents the full embedding structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic Contrastive Learning for Hierarchical 3D Point Cloud Embedding." pith.science (2026). https://pith.science/paper/RGQ5RTUU

@misc{pith2026250102285,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic Contrastive Learning for Hierarchical 3D Point Cloud Embedding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RGQ5RTUU}},
  note         = {Machine review of arXiv:2501.02285}
}
read the original abstract

Hyperbolic spaces allow for more efficient modeling of complex, hierarchical structures, which is particularly beneficial in tasks involving multi-modal data. Although hyperbolic geometries have been proven effective for language-image pre-training, their capabilities to unify language, image, and 3D Point Cloud modalities are under-explored. We extend the 3D Point Cloud modality in hyperbolic multi-modal contrastive pre-training. Additionally, we explore the entailment, modality gap, and alignment regularizers for learning hierarchical 3D embeddings and facilitating the transfer of knowledge from both Text and Image modalities. These regularizers enable the learning of intra-modal hierarchy within each modality and inter-modal hierarchy across text, 2D images, and 3D Point Clouds. Experimental results demonstrate that our proposed training strategy yields an outstanding 3D Point Cloud encoder, and the obtained 3D Point Cloud hierarchical embeddings significantly improve performance on various downstream tasks.

Figures

Figures reproduced from arXiv: 2501.02285 by the authors.

Figure 1
Figure 1. Hyperbolicity coverage curves and distribution of [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 3
Figure 3. Disentangled analysis for our obtained 3D Point Cloud [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 45 canonical work pages

  1. [1]

    Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud un- derstanding

    Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud un- derstanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 9902–9912, 2022. 2

    work page 2022

  2. [2]

    Llm2vec: Large lan- guage models are secretly powerful text encoders

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large lan- guage models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961, 2024. 1

    arXiv 2024

  3. [3]

    Bronstein, and Jonathan J

    Edoardo Cetin, Benjamin Paul Chamberlain, Michael M. Bronstein, and Jonathan J. Hunt. Hyperbolic deep reinforcement learning. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023. 1

    work page 2023

  4. [4]

    Grounded 3d-llm with referent tokens

    Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370 , 2024. 2

    arXiv 2024

  5. [5]

    Embedding Geometries of Contrastive Language-Image Pre-Training

    Jason Chuan-Chih Chou and Nahid Alam. Embedding geome- tries of contrastive language-image pre-training. arXiv preprint arXiv:2409.13079, 2024. 2

    work page Pith review arXiv 2024

  6. [6]

    Hyperbolic image-text rep- resentations

    Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Shanmukha Ramakrishna Vedantam. Hyperbolic image-text rep- resentations. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors, In- ternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA...

    work page 2023

  7. [7]

    Christiane Fellbaum. Wordnet. In Theory and applications of ontology: computer applications, pages 231–243. Springer, 2010. 1

    work page 2010

  8. [8]

    Computing the Gromov hyperbolicity of a discrete metric space

    Herv ´e Fournier, Anas Ismail, and Antoine Vigneron. Computing the Gromov hyperbolicity of a discrete metric space. Information Processing Letters, 115(6-8):576–579, 2015. 5

    work page 2015

Show all 57 references
  1. [9]

    Hy- perbolic entailment cones for learning hierarchical embeddings

    Octavian-Eugen Ganea, Gary B ´ecigneul, and Thomas Hofmann. Hy- perbolic entailment cones for learning hierarchical embeddings. In Jennifer G. Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stock- holmsm¨assan, ...

    2018

  2. [10]

    Hyperbolic contrastive learning for visual representations be- yond objects

    Songwei Ge, Shlok Mishra, Simon Kornblith, Chun-Liang Li, and David Jacobs. Hyperbolic contrastive learning for visual representations be- yond objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 6840–6849, 2023. 1

    2023

  3. [11]

    Hyperbolic groups

    M Gromov. Hyperbolic groups. Essays in Group Theory, pages/Springer-Verlag, 1987. 4

    1987

  4. [12]

    Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following

    Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arX...

    2023 arXiv

  5. [13]

    3d-llm: Injecting the 3d world into large language models

    Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482–20494, 2023. 1

    2023

  6. [14]

    Clip2point: Transfer clip to point cloud classification with image-depth pre-training

    Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Ryn- son WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2215...

    2023

  7. [15]

    Zeroshape: Regression-based zero-shot shape reconstruction

    Zixuan Huang, Stefan Stojanov, Anh Thai, Varun Jampani, and James M Rehg. Zeroshape: Regression-based zero-shot shape reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10061–10071, 2024. 5

    2024

  8. [16]

    Openclip, July 2021

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. 2

    2021

  9. [17]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482–7491, 2018. 5

    2018

  10. [18]

    Oseledets, and Victor S

    Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan V . Oseledets, and Victor S. Lempitsky. Hyperbolic image embeddings. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020 , pages 6417–

    2020

  11. [19]

    Hype: Hyperbolic entailment filtering for underspecified images and texts

    Wonjae Kim, Sanghyuk Chun, Taekyung Kim, Dongyoon Han, and Sangdoo Yun. Hype: Hyperbolic entailment filtering for underspecified images and texts. arXiv preprint arXiv:2404.17507 , 2024. 1, 2

    2024 arXiv

  12. [20]

    Geodesic self-attention for 3d point clouds

    Zhengyu Li, Xuan Tang, Zihao Xu, Xihao Wang, Hui Yu, Mingsong Chen, et al. Geodesic self-attention for 3d point clouds. Advances in Neural Information Processing Systems , 35:6190–6203, 2022. 7

    2022

  13. [21]

    Application of hyperbolic space attention mechanisms in 3d point cloud classification

    Lamei Liu and Zhiqiang Liu. Application of hyperbolic space attention mechanisms in 3d point cloud classification. In 2024 6th International Conference on Communications, Information System and Computer Engineering (CISCE), pages 658–662. IEEE, 2024. 1, 2

    2024

  14. [22]

    Hyperbolic learning with multimodal large language models

    Paolo Mandica, Luca Franco, Konstantinos Kallidromitis, Suzanne Petryk, and Fabio Galasso. Hyperbolic learning with multimodal large language models. arXiv preprint arXiv:2408.05097 , 2024. 1

    2024 arXiv

  15. [23]

    Rethinking the compositionality of point clouds through regularization in the hyperbolic space

    Antonio Montanaro, Diego Valsesia, and Enrico Magli. Rethinking the compositionality of point clouds through regularization in the hyperbolic space. Advances in Neural Information Processing Systems , 35:33741– 33753, 2022. 1, 2, 6

    2022

  16. [24]

    Towards hyperbolic regularizers for point cloud part segmentation

    Antonio Montanaro, Diego Valsesia, and Enrico Magli. Towards hyperbolic regularizers for point cloud part segmentation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5. IEEE, 2023. 1, 2, 7

    2023

  17. [25]

    Poincar ´e embeddings for learning hierarchical representations

    Maximilian Nickel and Douwe Kiela. Poincar ´e embeddings for learning hierarchical representations. CoRR, abs/1705.08039, 2017. 1, 2

    2017 arXiv

  18. [26]

    A text is worth several tokens: Text embedding from llms secretly aligns well with the key tokens

    Zhijie Nie, Richong Zhang, and Zhanyu Wu. A text is worth several tokens: Text embedding from llms secretly aligns well with the key tokens. arXiv preprint arXiv:2406.17378 , 2024. 1

    2024 arXiv

  19. [27]

    Compositional entailment learning for hyperbolic vision-language models

    Avik Pal, Max van Spengler, Guido Maria D’Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. arXiv preprint arXiv:2410.06912, 2024. 2

    arXiv

  20. [28]

    Hierarchical structure in perceptual representation

    Stephen E Palmer. Hierarchical structure in perceptual representation. Cognitive psychology, 9(4):441–474, 1977. 1

    1977

  21. [29]

    Masked autoencoders for point cloud self-supervised learning

    Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. arXiv preprint arXiv:2203.06604 , 2022. 3

    2022 arXiv

  22. [30]

    Development of fast algorithms for calcu- lating delta-hyperbolicity of data in collaborative filtering problems in python

    Matveeva Tatana Pavlovna. Development of fast algorithms for calcu- lating delta-hyperbolicity of data in collaborative filtering problems in python. DSpace at Saint Petersburg State University , 2024. 5

    2024

  23. [31]

    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, 2017. 7

    2017

  24. [32]

    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. 7

    2017

  25. [33]

    Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining

    Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. In International Conference on Machine Learning, pages 28223–28243. PMLR, 2023. 2, 3, 5, 7

    2023

  26. [34]

    Shapellm: Universal 3d object 9 understanding for embodied interaction

    Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object 9 understanding for embodied interaction. In European Conference on Computer Vision, pages 214–238. Springer, 2025. 1, 2, 4

    2025

  27. [35]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning , p...

    2021

  28. [36]

    Improving dictionary learning with gated sparse autoencoders

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, J´anos Kram´ar, Rohin Shah, and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. arXiv preprint arXiv:2404.16014, 2024. 7

    2024 arXiv

  29. [37]

    Accept the modality gap: An exploration in the hyperbolic space

    Sameera Ramasinghe, Violetta Shevchenko, Gil Avraham, and Ajanthan Thalaiyasingam. Accept the modality gap: An exploration in the hyperbolic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 27263–27272, 2024. 1, 2, 3, 4, 6

    2024

  30. [38]

    The role of hierarchical knowledge representation in decisionmaking and system management

    Jens Rasmussen. The role of hierarchical knowledge representation in decisionmaking and system management. IEEE Transactions on systems, man, and cybernetics , (2):234–243, 1985. 1

    1985

  31. [39]

    Hyperbolic neural networks++

    Ryohei Shimizu, YUSUKE Mukuta, and Tatsuya Harada. Hyperbolic neural networks++. In International Conference on Learning Represen- tations, 2021. 1, 2

    2021

  32. [40]

    Representation learning with contrastive predictive coding

    A ¨aron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748,

    arXiv

  33. [41]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V . N. Vishwanathan, and Roman Garnett,...

    2017

  34. [42]

    Hypliloc: Towards effective lidar pose regression with hyperbolic fusion

    Sijie Wang, Qiyu Kang, Rui She, Wei Wang, Kai Zhao, Yang Song, and Wee Peng Tay. Hypliloc: Towards effective lidar pose regression with hyperbolic fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5176–5185, 2023. 2

    2023

  35. [43]

    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,

  36. [44]

    Gonzalez, and Peter Vajda

    Bichen Wu, Ruizhe Cheng, Peizhao Zhang, Tianren Gao, Joseph E. Gonzalez, and Peter Vajda. Data efficient language-supervised zero-shot recognition with optimal transport distillation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, ...

    2022

  37. [45]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang 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 Recognition , pages 1912–1920, 2015. 6, 7

    1912

  38. [46]

    Hecpg: Hyperbolic embedding and confident patch-guided network for point cloud matching

    Yifan Xie, Jihua Zhu, Shiqi Li, Naiwen Hu, and Pengcheng Shi. Hecpg: Hyperbolic embedding and confident patch-guided network for point cloud matching. IEEE Transactions on Geoscience and Remote Sensing,

  39. [47]

    Revealing the dark secrets of masked image modeling

    Zhenda Xie, Zigang Geng, Jingcheng Hu, Zheng Zhang, Han Hu, and Yue Cao. Revealing the dark secrets of masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14475–14485, 2023. 3

    2023

  40. [48]

    Enhancing llm complex reasoning capability through hyperbolic geometry

    Menglin Yang, Aosong Feng, Bo Xiong, Jiahong Liu, Irwin King, and Rex Ying. Enhancing llm complex reasoning capability through hyperbolic geometry. In ICML 2024 Workshop on LLMs and Cognition ,

    2024

  41. [49]

    A scalable active framework for region annotation in 3d shape collections

    Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (TOG) , 35(6):1–12, 2016. 7

    2016

  42. [50]

    Coca: Contrastive captioners are image-text foundation models

    Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyed- hosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Trans. Mach. Learn. Res. , 2022, 2022. 2

    2022

  43. [51]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. arXiv preprint arXiv:2111.14819 , 2021. 7

    2021 arXiv

  44. [52]

    Hyperbolic contrastive learning

    Yun Yue, Fangzhou Lin, Kazunori D Yamada, and Ziming Zhang. Hyperbolic contrastive learning. arXiv preprint arXiv:2302.01409, 2023. 1

    2023 arXiv

  45. [53]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF Inter- national Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pages 11941–11952. IEEE, 2023. 2

    2023

  46. [54]

    Pointclip: Point cloud understanding by clip

    Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8552–8562, 2022. 2

    2022

  47. [55]

    Manning, and Curtis P

    Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive learning of medical visual rep- resentations from paired images and text. In Zachary C. Lipton, Rajesh Ranganath, Mark P. Sendak, Michael W. Sjoding, and Serena Yeung, editors,...

    2022

  48. [56]

    Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning

    Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 2639– 2650, 2023. 2

    2023

  49. [6427]

    Computer Vision Foundation / IEEE, 2020. 1, 2, 5

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.