Pith. sign in

REVIEW 3 major objections 6 minor 159 references

PointCHR: Point Cloud Analysis via Curvature-Aware Hyperbolic Rectification

T0 review · 3 major / 6 minor · reviewed 2026-07-31 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a learnable curvature-aware radial dilation on the Poincaré ball un-crowds high-curvature point-cloud features, yielding state-of-the-art results on segmentation and classification benchmarks.

desk verdict The plug-in's gains look real and the curvature-stratified evidence is genuinely supportive, but the hyperbolic-capacity mechanism needs a tighter ablation before the SOTA claims are taken at face value. read the letter →

arxiv 2607.24052 v1 pith:4NH3V7QD submitted 2026-07-27 cs.CV

Xinxing Yu , Liying Yang , Hao Mo , Hui Ma , Fang Kai , Ajian Liu , Yanyan Liang This is my paper
classification cs.CV
keywords pointcloudanalysishyperbolicgeometryPoincaréballcurvatureestimationrepresentationcrowdingMöbiustransformationsemanticsegmentationshapeclassification
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

PointCHR argues that high-curvature regions of a 3D point cloud—corners, edges, fine details—are sparse but informative, and that Euclidean feature spaces, with only polynomial volume growth, crowd them together with smooth surfaces. The core move is to route point features through a Poincaré ball, where volume grows exponentially with distance from the origin, and to learn a per-point scale that pushes high-curvature points outward while leaving flat points near the center. A closed-form Möbius dilation implements this without the gradient and overflow problems of operating near the boundary. The authors report state-of-the-art accuracy on semantic segmentation, part segmentation, and classification benchmarks, with the largest gains in the highest-curvature bins. A sympathetic reader would take the paper to be establishing that geometry-aware radial repartitioning of embedding capacity—rather than more parameters—is what fixes fine-grained point-cloud discrimination.

What carries the argument

The CHR block has three stages. HST lifts Euclidean features to the Poincaré ball via the exponential map and recombines them with Möbius linear layers and Möbius GELU. PCP computes a softplus-gated, curvature-conditioned scale factor s_i = 1 + α·g_i·(κ_i)^γ using an MLP that fuses semantic features with PCA-estimated curvature. CGD applies the closed-form Möbius scalar multiplication, Eq. (15), to multiply each point's geodesic distance from the origin by s_i; because t→tanh(t) maps into (−1,1), the dilated embedding stays in the ball without tangent-space projection, avoiding gradient vanishing near the boundary. The mechanism's job is to convert the curvature long-tail into a radial order

What would settle it

Run S3DIS with an otherwise identical model in which the Möbius dilation of Eq. (15) is replaced by a Euclidean radial rescaling that maps the learned scale s_i onto the same distribution of embedding norms, then compare tail-bin mIoU and classification margins. If the Euclidean version matches or nearly matches PointCHR, the claimed exponential-capacity advantage is not load-bearing; the margin and ED differences reported in the paper should shrink toward zero.

Watch

Extended reading notes

Core claim

Central claim: sparse high-curvature points are poorly served by Euclidean capacity, which grows only polynomially, so PointCHR learns a per-point hierarchy scale s_i = 1 + α·g_i·(κ_i)^γ from curvature and local context, then applies it as a geodesic dilation on the Poincaré ball. The dilation is computed in closed form by the Möbius scalar multiplication z_i^(2) = (1/√c) tanh(s_i · arctanh(√c·r̃_i)) z_i / r̃_i, which scales geodesic distance from the origin while preserving direction and automatically keeps the point inside the ball. High-curvature points are therefore moved toward the high-capacity boundary and low-curvature points stay near the stable origin. The paper reports that this r

Load-bearing premise

The load-bearing premise is that distance from the origin in the Poincaré ball reliably buys usable representational capacity for point-cloud semantics; if radial position does not confer useful capacity, the method reduces to anisotropic scaling with extra parameters and the hyperbolic narrative loses support.

Editorial extensions

If this is right

  • If the radial-capacity claim is correct, networks can allocate representation to geometrically complex regions without scaling up parameters; PointCHR reports S3DIS mIoU 86.0 with 21.0M parameters and 4.0 GFLOPs, a modest increase over its base.
  • The rectification is add-on and architecture-agnostic: the authors report consistent gains from adding CHR to PointMLP, DeLA, and PointNeXt-s backbones.
  • The mechanism's signature is a tail-uplift: on S3DIS the highest-curvature bin rises by 10.40 mIoU, while lower-curvature bins gain less, and boundary points receive larger learned scales than interior points.
  • Euclidean alternatives are not equivalent: the appendix reports that anisotropic Euclidean scaling helps but trails the hyperbolic version, and curvature-guided resampling hurts, supporting the paper's claim that the bottleneck is feature-space capacity rather than point density.
  • The geometry prior is not brittle: per the robustness experiments, S3DIS mIoU stays above 85 under moderate curvature noise and stays near 86 until 30% masking of points or curvature values.

Reading between the lines

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

  • Editorial extension: if the radial-capacity mechanism is what carries the gain, the same dilation should work when driven by any sparse per-point geometric signal—surface normals, boundary proximity, or local shape variance—not only PCA curvature; this is directly testable.
  • Editorial extension: the appendix's observation that medium-curvature transitions gain least is a concrete prediction of the radial-capacity story: those points sit at intermediate radii where the conformal factor grows slowly. A synthetic surface experiment with controlled curvature could verify the predicted nonlinear gain curve.
  • Editorial extension: the cleanest necessity test is to keep the learned scaling schedule fixed and replace Eq. (15) with a Euclidean rescaling that reproduces the same output norm distribution. If the S3DIS gap closes, the load-bearing part is the monotone radial renormalization, not hyperbolic geometry per se.
  • Editorial extension: since gains on clean global classification (ModelNet40) are modest, the practical payoff is concentrated in dense scene and part understanding, where local boundaries dominate the error.
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

3 major / 6 minor

Summary. The paper proposes PointCHR, a plug-in module that maps point-cloud features into the Poincaré ball, applies a learnable curvature-aware radial dilation, and maps back to Euclidean space for downstream tasks. After an encoder, point-wise PCA curvature κ is combined with semantic features through a gating MLP to produce a per-point scale s_i = 1 + α·g_i·κ_i^γ (Eq. 13). The module then applies Möbius linear/GELU layers, a closed-form geodesic dilation z^(2)_i = (1/√c) tanh(s_i·arctanh(√c·r̃_i)) z^(1)_i / r̃_i (Eq. 15), and a log-map back to Euclidean space. The authors report state-of-the-art results on S3DIS (86.0/89.1 mIoU), ShapeNetPart (87.0/85.7), ScanObjectNN (92.7/91.7), and ModelNet40 (93.7/92.0), with ablations, a Euclidean anisotropic scaling baseline, robustness tests, and curvature-stratified analyses. The central claimed mechanism is that hyperbolic exponential volume growth near the boundary 'unfolds' crowded high-curvature representations.

Significance. If the mechanism is correct, the module is a simple, model-agnostic component that improves boundary-heavy segmentation and classification, with respectable complexity overhead (21.0M params, 4.0 GFLOPs). The closed-form CGD is mathematically sound and the experimental protocol is transparent (single-view evaluation, standard splits). The paper also includes several good controls: a Euclidean anisotropic scaling (AES) baseline, a curvature-guided resampling baseline, sensitivity analyses over neighborhood size, curvature parameters, noise, and missing-curvature robustness. These are genuine strengths. However, the load-bearing claim that the gains come from exponential hyperbolic radial capacity rather than from anisotropic scaling plus extra trainable components is not yet established by the evidence presented. The AES ablation is not a controlled comparison, and the Gromov-δ and ED/margin measurements are insufficiently specified. Since the paper's novelty is precisely the hyperbolic-capacity explanation, this gap is substantive rather than cosmetic.

major comments (3)
  1. [Appendix C, Table 12; §4.2] The central causal claim that hyperbolic exponential volume growth, not generic anisotropic scaling, drives the improvement is not supported by the provided ablation. The AES variant removes the entire HST pipeline—Möbius linear/GELU layers, exp/log maps, and the manifold structure—so the residual 1.08 mIoU (S3DIS) / 0.35 OA (ScanObjectNN) confounds geometry with additional parameters and optimization paths. Please add a controlled Euclidean twin that keeps the full HST pipeline (Euclidean linear+GELU, the same radial scaling s_i, and the same log/exp-style feature mapping) and only replaces the Poincaré metric with Euclidean distance scaling. Alternatively, apply Eqs. (13)–(15) but with Euclidean radial dilation inside the HST pipeline. Without this control, the abstract's statement that the gain is due to 'larger effective embedding capacities' near the hyperbolic boundary is not empir
  2. [Appendix D, §5.8, Tables 11/13/14] The Gromov δ ≈ 0.02 claim is asserted without any measurement details (which substructures, which scale, how δ was computed), so it cannot independently justify the near-tree hyperbolic assumption. The ED and margin metrics are self-defined, and Table 11 shows mixed effects (ED decreases in low-curvature bins while increasing in high-curvature bins), so their interpretation as 'representation crowding is alleviated' needs a more direct test. Please provide (a) the exact procedure for estimating δ with error bars over scenes/objects, and (b) a controlled comparison of feature separation under the learned s_i versus a fixed uniform scale s=1 or a random permutation of the same s values, measured both by ED/margin and by downstream stratified performance.
  3. [§5.2, Tables 1, 3, 4, 5; §5.3, Table 6] All headline results are from single optimization runs, with no standard deviations or significance tests. Several reported gains are small relative to typical run-to-run variation in point-cloud benchmarks (e.g., ModelNet40 OA 93.7 ties PointSD's 93.7; ScanObjectNN OA 92.1→92.7 is a 0.6-point gain). The text repeatedly uses 'significant improvements' and 'state-of-the-art' without statistical support. Please report means and standard deviations over at least three seeds for the core comparisons, or otherwise justify that the differences are outside run-to-run noise. This is necessary to support the quantitative SOTA claims, especially because the method's benefits vary by task and backbone.
minor comments (6)
  1. [§3.2, Eq. (3)] The definition of the closed ball B^d_{c,ε} is notationally ambiguous: the text reads '||x|| ≤ 1−ε√c' but the intended bound is presumably ||x|| ≤ (1−ε)/√c. Please correct the parentheses.
  2. [§3.2, Eq. (5)] The log map expression divides by ||x||; the case x=0 should be handled explicitly (e.g., by returning 0). The current notation leaves this implicit.
  3. [§4.2, Eq. (11)] The 'Möbius variant of the GELU activation' σ_{⊗_c} is not defined. A pointwise GELU in the Poincaré ball is not automatically well-defined, so please give the explicit formula (e.g., Möbius scalar multiplication with the GELU nonlinearity applied in the tangent space or via a standard hyperbolic activation) or cite the exact construction used.
  4. [§5.3, Table 6] The checkmark column layout of Table 6 is confusing: the columns are not clearly labeled, and the rows are inconsistent with the text's description of which component (HST alone vs PCP alone) yields the larger gain. Please reformat the table so each row unambiguously indicates which components are active.
  5. [Figure 1 caption] Typo: 'datast' should be 'dataset'.
  6. [Appendix F.4, Table 19] The curvature noise σ=0.01 row shows an improvement (+0.36 mIoU) over the clean baseline, which is not explained. A brief remark on why small noise can be beneficial would help interpretation.

Circularity Check

0 steps flagged · score 1.0 of 10

No equation-level circularity: the SOTA numbers are external benchmarks, and the radial-scaling design is a construction, not a fitted prediction.

full rationale

The core results (S3DIS 86.0/89.1 mIoU, ShapeNetPart 87.0/85.7, ScanObjectNN 92.7/91.7, ModelNet40 93.7/92.0) are measured against standard external benchmarks, so they cannot reduce to the paper's own equations. The CHR block is a fixed architecture: HST uses standard exponential/Möbius maps; PCP computes s_i = 1 + α g_i κ_i^γ (Eq. 13) with fixed α, γ; and CGD applies the closed-form Möbius scalar multiplication (Eq. 15). No parameter is fitted to the reported test numbers and then renamed a prediction. The observation that high-curvature points receive larger scaling factors (Table 10) is partly a consequence of Eq. 13, but g_i is a learned, data-dependent gate and the actual mIoU/ED/margin changes are empirical. The 'representation crowding' construct is defined in Appendix D via the authors' own ED/margin metrics, which is somewhat self-referential, but those metrics are measured post hoc and the headline results do not depend on that definition. Appendix C's AES ablation is not a clean control because it omits HST, and the Gromov δ = 0.02 is asserted without measurement details; these are evidence-quality concerns, not circularity under the definitions used here. No load-bearing self-citation chain, imported uniqueness theorem, or ansatz smuggled via citation appears. Score 1 reflects only the mild self-referential flavor of the mechanistic narrative, not an equation-level circularity.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The paper contributes a heuristic mechanism whose success is carried by a standard hyperbolic-geometry toolkit, a hand-chosen scaling law (Eq. 13), and a learned gating MLP. The free parameters are few and sensitivity is reported, but the core 'capacity' premise is assumed rather than derived. No new physical or mathematical entities are introduced; 'representation crowding' is a framing device, not an entity.

free parameters (8)
  • Dilation amplitude α = 1.0
    Eq. 13; hand-set maximum radial dilation; sensitivity table shows mild variation over {1,2,3}.
  • Curvature sensitivity γ = 2.0
    Eq. 13; hand-set exponent controlling how strongly κ affects the scaling factor; tested over 1,2,3.
  • Manifold curvature c = 1.0
    Controls the Poincaré ball curvature; hand-set and fixed; sensitivity tested at 0.5, 1.0, 2.0.
  • Stability margin ε = 1e-5
    Numerical margin in exp/log maps and CGD, Eq. 3 and Algorithm 1; affects boundary handling.
  • Neighborhood size k = 24
    Default k for PCA curvature estimation, Eq. 1; sensitivity tested at 16, 24, 32.
  • Curvature clipping threshold κmax = 0.2
    Appendix G; used in min-max normalization and clipping of κ to suppress outliers.
  • Boundary edge radius R = 0.02
    Evaluation-only threshold for edge-strip segmentation, Table 2; not a model parameter but affects reported edge results.
  • Gating MLP Gφ weights = learned
    Eq. 12; the softplus MLP fusing h_i and κ_i is trained end-to-end; its learned weights are part of the scaling mechanism.
assumptions (5)
  • domain assumption Poincaré ball's exponential volume growth translates into larger useful embedding capacity near the boundary.
    Used throughout Section 4.2 to justify radial dilation; the geometric fact is standard, but the capacity-for-semantics step is assumed, not proven.
  • domain assumption PCA-based local curvature κ_i (Eq. 2) is a reliable proxy for informative geometric complexity.
    The PCP mechanism gates the scaling factor on κ; robustness tests in Appendix F support stability but not semantic validity.
  • standard math Euclidean features can be identified with the tangent space at the origin of the Poincaré ball and mapped losslessly via the exponential map.
    Eq. 9; standard differential-geometric identification, but it treats the backbone's Euclidean feature norm as meaningful.
  • domain assumption Möbius linear layer and Möbius GELU preserve the relevant hyperbolic geometry for point-cloud features.
    Imported from Ganea et al. and applied without data-specific verification.
  • ad hoc to paper Point cloud manifolds are δ-hyperbolic with δ ≈ 0.02 (near-tree).
    Appendix D asserts this measurement without a described procedure; it is used to motivate why Euclidean embeddings struggle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PointCHR: Point Cloud Analysis via Curvature-Aware Hyperbolic Rectification." pith.science (2026). https://pith.science/paper/4NH3V7QD

@misc{pith2026260724052,
  author       = {Pith},
  title        = {Pith review of: PointCHR: Point Cloud Analysis via Curvature-Aware Hyperbolic Rectification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4NH3V7QD}},
  note         = {Machine review of arXiv:2607.24052}
}
read the original abstract

High-curvature regions in 3D point clouds encapsulate critical fine-grained geometric semantics yet exhibit a distinct long-tail sparsity in their spatial distribution. The inherent limitations of polynomial volume growth in Euclidean space frequently render these intricate geometric features challenging to adequately resolve within a uniform-scale feature space. Consequently, these regions are frequently overshadowed by smooth global features dominated by low-curvature regions, thereby limiting the discriminative capacity of the network. To address this issue, we propose PointCHR, a curvature-aware hyperbolic rectification (CHR) for point cloud analysis. Utilising the property of exponential volume expansion in the vicinity of hyperbolic manifolds, CHR presents a learnable curvature-guided radial rectification mechanism. By adaptively projecting high-curvature points towards boundary regions endowed with larger effective embedding capacities, PointCHR effectively mitigates the representation crowding problem inherent in Euclidean settings. Extensive experimentation has demonstrated that PointCHR significantly enhances the ability of backbone to capture fine-grained geometric details, achieving state-of-the-art performance across multiple benchmarks.

Figures

Figures reproduced from arXiv: 2607.24052 by the authors.

Figure 1
Figure 1. Illustration of the curvature imbalance challenge. (a) Quantitative analysis shows the long-tailed nature of curvature distribution in S3DIS datast. (b) In complex corner regions, con￾ventional encoders fail to handle such tail samples, leading to boundary artifacts and noise. Our CHR addresses this by incor￾porating geometric priors for feature optimization, successfully restoring crisp boundaries consistent with t… view at source ↗
Figure 2
Figure 2. (a) Overview of the proposed PointCHR framework. (b) The Euclidean baseline struggles with this data imbalance, showing a sharp performance degradation in the high-curvature tail. (c) Our PointCHR effectively rectifies this issue. By leveraging hyperbolic geometry to handle structural complexity, our method prevents performance collapse in the tail, demonstrating superior robustness on rare hard samples. The backgro… view at source ↗
Figure 3
Figure 3. Qualitative comparison of segmentation results on the S3DIS Area5 dataset. PointCHR effectively preserves fine-grained boundary details and semantic consistency, significantly outperforming the baseline in challenging scenarios [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

159 extracted references · 1 canonical work pages

  1. [1]

    , title =

    Guo, Yunhui and Wang, Xudong and Chen, Yubei and Yu, Stella X. , title =. CVPR , month =. 2022 , pages =

    2022

  2. [2]

    CVPR , pages=

    Hyperbolic vision transformers: Combining improvements in metric learning , author=. CVPR , pages=

  3. [3]

    CVPR , pages=

    Hyperbolic image embeddings , author=. CVPR , pages=

  4. [4]

    CVPR , pages=

    Hyperbolic visual embedding learning for zero-shot recognition , author=. CVPR , pages=

  5. [5]

    NeurIPS , volume=

    Hyperbolic graph convolutional neural networks , author=. NeurIPS , volume=

  6. [6]

    Starting from Non-Parametric Networks for 3D Point Cloud Analysis , year=

    Zhang, Renrui and Wang, Liuhui and Wang, Yali and Gao, Peng and Li, Hongsheng and Shi, Jianbo , booktitle=. Starting from Non-Parametric Networks for 3D Point Cloud Analysis , year=

  7. [7]

    CVPR , pages=

    Relation-shape convolutional neural network for point cloud analysis , author=. CVPR , pages=

  8. [8]

    CVPR , month =

    Tang, Liyao and Zhan, Yibing and Chen, Zhe and Yu, Baosheng and Tao, Dacheng , title =. CVPR , month =. 2022 , pages =

    2022

Show all 159 references
  1. [9]

    ACM Transactions on Graphics (TOG) , volume=

    Deltaconv: anisotropic operators for geometric deep learning on point clouds , author=. ACM Transactions on Graphics (TOG) , volume=. 2022 , publisher=

    2022

  2. [10]

    AAAI , pages =

    Mutian Xu and Junhao Zhang and Zhipeng Zhou and Mingye Xu and Xiaojuan Qi and Yu Qiao , title =. AAAI , pages =

  3. [11]

    CVPR , month =

    Qin, Zheng and Yu, Hao and Wang, Changjian and Guo, Yulan and Peng, Yuxing and Xu, Kai , title =. CVPR , month =. 2022 , pages =

    2022

  4. [12]

    ICCV , month =

    Xiang, Tiange and Zhang, Chaoyi and Song, Yang and Yu, Jianhui and Cai, Weidong , title =. ICCV , month =. 2021 , pages =

    2021

  5. [13]

    CVPR , month =

    Ran, Haoxi and Liu, Jun and Wang, Chengjie , title =. CVPR , month =. 2022 , pages =

    2022

  6. [14]

    ICML , year=

    Exploring Vision Semantic Prompt for Efficient Point Cloud Understanding , author=. ICML , year=

  7. [15]

    NeurIPS , year=

    ZigzagPointMamba: Spatial-Semantic Mamba for Point Cloud Understanding , author=. NeurIPS , year=

  8. [16]

    ICCV , pages=

    Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation , author=. ICCV , pages=

  9. [17]

    CVPR , pages=

    Pointconv: Deep convolutional networks on 3d point clouds , author=. CVPR , pages=

  10. [18]

    IROS , year=

    Real-time Dexterous Telemanipulation with an End-Effect-Oriented Learning-based Approach , author=. IROS , year=

  11. [19]

    NeurIPS , year=

    PointMapPolicy: Structured Point Cloud Processing for Multi-Modal Imitation Learning , author=. NeurIPS , year=

  12. [20]

    ECCV , pages=

    Rapid-seg: Range-aware pointwise distance distribution networks for 3d lidar segmentation , author=. ECCV , pages=. 2024 , organization=

    2024

  13. [21]

    ICCV , pages=

    On offline evaluation of 3d object detection for autonomous driving , author=. ICCV , pages=

  14. [22]

    ICCV , month =

    Liu, Linshen and Su, Boyan and Jiang, Junyue and Wu, Guanlin and Guo, Cong and Xu, Ceyu and Yang, Hao Frank , title =. ICCV , month =. 2025 , pages =

    2025

  15. [23]

    ICLR , year =

    Ilya Loshchilov and Frank Hutter , title =. ICLR , year =

  16. [24]

    Hyperbolic Deep Neural Networks: A Survey , year=

    Peng, Wei and Varanka, Tuomas and Mostafa, Abdelrahman and Shi, Henglin and Zhao, Guoying , journal=. Hyperbolic Deep Neural Networks: A Survey , year=

  17. [25]

    Nickel, Maximillian and Kiela, Douwe , journal=. Poincar

  18. [26]

    arXiv preprint arXiv:1606.08415 , year=

    Gaussian Error Linear Units (Gelus) , author=. arXiv preprint arXiv:1606.08415 , year=

    arXiv

  19. [27]

    2008 , publisher=

    Analytic hyperbolic geometry and Albert Einstein's special theory of relativity , author=. 2008 , publisher=

    2008

  20. [28]

    NeurIPS , volume=

    Hyperbolic neural networks , author=. NeurIPS , volume=

  21. [29]

    , author=

    Analysis of a complex of statistical variables into principal components. , author=. Journal of educational psychology , volume=

  22. [30]

    and Gross, M

    Pauly, M. and Gross, M. and Kobbelt, L.P. , booktitle=. Efficient simplification of point-sampled surfaces , year=

  23. [31]

    Journal of machine learning research , volume=

    Visualizing data using t-SNE , author=. Journal of machine learning research , volume=

  24. [32]

    2025 , booktitle =

    Li, Zhuoyuan and Ai, Yubo and Lu, Jiahao and Wang, ChuXin and Deng, Jiacheng and Chang, Hanzhi and Liang, Yanzhe and Yang, Wenfei and Zhang, Shifeng and Zhang, Tianzhu , title =. 2025 , booktitle =

    2025

  25. [33]

    AAAI , year=

    FEAST-Mamba: FEAture and SpaTial Aware Mamba Network with Bidirectional Orthogonal Fusion for Cross-Modal Point Cloud Segmentation , author=. AAAI , year=

  26. [34]

    Transformers are

    Dao, Tri and Gu, Albert , booktitle=. Transformers are

  27. [35]

    ACMMM , pages=

    Mambamos: Lidar-based 3d moving object segmentation with motion-aware state space model , author=. ACMMM , pages=

  28. [36]

    ACMMM , pages=

    Point-gcc: Universal self-supervised 3d scene pre-training via geometry-color contrast , author=. ACMMM , pages=

  29. [37]

    CVPR , pages=

    Mitigating object dependencies: Improving point cloud self-supervised learning through object exchange , author=. CVPR , pages=

  30. [38]

    CVPR , pages=

    Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding , author=. CVPR , pages=

  31. [39]

    IEEE Robotics and Automation Letters , volume=

    Segcontrast: 3d point cloud feature representation learning through self-supervised segment discrimination , author=. IEEE Robotics and Automation Letters , volume=. 2022 , publisher=

    2022

  32. [40]

    CVPR , pages=

    Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding , author=. CVPR , pages=

  33. [41]

    ECCV , pages=

    Shape2scene: 3d scene representation learning through pre-training on shape data , author=. ECCV , pages=. 2024 , organization=

    2024

  34. [42]

    CVPR , pages=

    Mm-3dscene: 3d scene understanding by customizing masked modeling with informative-preserved reconstruction and self-distilled consistency , author=. CVPR , pages=

  35. [43]

    ICCV , pages=

    Ponder: Point cloud pre-training via neural rendering , author=. ICCV , pages=

  36. [44]

    CVPR , pages=

    Masked scene contrast: A scalable framework for unsupervised 3d representation learning , author=. CVPR , pages=

  37. [45]

    ECCV , pages=

    A closer look at invariances in self-supervised pre-training for 3d vision , author=. ECCV , pages=. 2022 , organization=

    2022

  38. [46]

    CVPR , pages=

    Exploring data-efficient 3d scene understanding with contrastive scene contexts , author=. CVPR , pages=

  39. [47]

    ICLR , year=

    Gated Delta Networks: Improving Mamba2 with Delta Rule , author=. ICLR , year=

  40. [48]

    ICCV , pages=

    Emerging properties in self-supervised vision transformers , author=. ICCV , pages=

  41. [49]

    CVPR , pages=

    Self-supervised pre-training with masked shape prediction for 3d scene understanding , author=. CVPR , pages=

  42. [50]

    ICCV , pages=

    Spatio-temporal self-supervised representation learning for 3d point clouds , author=. ICCV , pages=

  43. [51]

    ECCV , pages=

    Pointcontrast: Unsupervised pre-training for 3d point cloud understanding , author=. ECCV , pages=. 2020 , organization=

    2020

  44. [52]

    NeurIPS , volume=

    Point cloud matters: Rethinking the impact of different observation spaces on robot learning , author=. NeurIPS , volume=

  45. [53]

    CVPR , pages=

    PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipulation , author=. CVPR , pages=

  46. [54]

    AAAI , volume=

    L4dr: Lidar-4dradar fusion for weather-robust 3d object detection , author=. AAAI , volume=

  47. [55]

    CVPR , pages=

    Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection , author=. CVPR , pages=

  48. [56]

    PAMI , volume=

    Exploring point-BEV fusion for 3D point cloud object tracking with transformer , author=. PAMI , volume=. 2024 , publisher=

    2024

  49. [57]

    CVPR , pages=

    Towards robust 3d object detection with lidar and 4d radar fusion in various weather conditions , author=. CVPR , pages=

  50. [58]

    CVPR , month=

    Open3DSG: Open-Vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-Set Relationships , author=. CVPR , month=

  51. [59]

    CVPR , year =

    Wang, Yan and Jia, Baoxiong and Zhu, Ziyu and Huang, Siyuan , title =. CVPR , year =

  52. [60]

    CVPR , month =

    Zhu, Liyuan and Huang, Shengyu and Schindler, Konrad and Armeni, Iro , title =. CVPR , month =. 2024 , pages =

    2024

  53. [61]

    NeurIPS , volume=

    A unified framework for 3d scene understanding , author=. NeurIPS , volume=

  54. [62]

    CVPR , year=

    KPConvX: Modernizing Kernel Point Convolution with Kernel Attention , author=. CVPR , year=

  55. [63]

    CVPR , year=

    3d semantic parsing of large-scale indoor spaces , author=. CVPR , year=

  56. [64]

    CVPR , year =

    ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes , author=. CVPR , year =

  57. [65]

    AAAI , author =

    Point. AAAI , author =. 2025 , pages =

    2025

  58. [66]

    IROS , author =

    Voxnet:. IROS , author =. 2015 , pages =

    2015

  59. [67]

    NeurIPS , author =

    Linnet:. NeurIPS , author =. 2024 , pages =

    2024

  60. [68]

    Han, Xu and Tang, Yuan and Wang, Zhaoxuan and Li, Xianzhi , month = oct, year =

  61. [69]

    IEEE Transactions on Multimedia , author =

    Pointgl:. IEEE Transactions on Multimedia , author =. 2024 , pages =

    2024

  62. [70]

    IEEE Transactions on Intelligent Transportation Systems , author =

    Mssf:. IEEE Transactions on Intelligent Transportation Systems , author =

  63. [71]

    ACM Computing Surveys , author =

    Point. ACM Computing Surveys , author =. 2025 , pages =. doi:10.1145/3715851 , language =

    2025 doi

  64. [72]

    , year =

    Park, Jinyoung and Lee, Sanghyeok and Kim, Sihyeon and Xiong, Yunyang and Kim, Hyunwoo J. , year =. Self-positioning point-based transformer for point cloud understanding , urldate =

  65. [73]

    and Su, Hao and Mo, Kaichun and Guibas, Leonidas J

    Qi, Charles R. and Su, Hao and Mo, Kaichun and Guibas, Leonidas J. , year =. Pointnet:

  66. [74]

    An end-to-end robust point cloud semantic segmentation network with single-step conditional diffusion models , urldate =

    Qu, Wentao and Wang, Jing and Gong, YongShun and Huang, Xiaoshui and Xiao, Liang , year =. An end-to-end robust point cloud semantic segmentation network with single-step conditional diffusion models , urldate =

  67. [75]

    2024 , keywords =

    Neurocomputing , author =. 2024 , keywords =

    2024

  68. [76]

    and Deschaud, Jean-Emmanuel and Marcotegui, Beatriz and Goulette, François and Guibas, Leonidas J

    Thomas, Hugues and Qi, Charles R. and Deschaud, Jean-Emmanuel and Marcotegui, Beatriz and Goulette, François and Guibas, Leonidas J. , year =. Kpconv:

  69. [77]

    and Kaiser, Lukasz and Polosukhin, Illia , year =

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia , year =. Attention is all you need , volume =

  70. [78]

    Wang, Ao and Chen, Hui and Lin, Zijia and Han, Jungong and Ding, Guiguang , year =

  71. [79]

    Wang, Changshuo and He, Shuting and Fang, Xiang and Han, Jiawei and Liu, Zhonghang and Ning, Xin and Li, Weijun and Tiwari, Prayag , year =. Point

  72. [80]

    Wang, Weijia , year =. Real-

  73. [81]

    ACM Transactions on Graphics , author =

    Dynamic. ACM Transactions on Graphics , author =. 2019 , pages =

    2019

  74. [82]

    2024 , journal=

    Zicheng Wang and Zhenghao Chen and Yiming Wu and Zhen Zhao and Luping Zhou and Dong Xu , title=. 2024 , journal=

    2024

  75. [83]

    Point transformer v3:

    Wu, Xiaoyang and Jiang, Li and Wang, Peng-Shuai and Liu, Zhijian and Liu, Xihui and Qiao, Yu and Ouyang, Wanli and He, Tong and Zhao, Hengshuang , year =. Point transformer v3:

  76. [84]

    IEEE Transactions on Multimedia , author =

    Group multi-view transformer for 3d shape analysis with spatial encoding , urldate =. IEEE Transactions on Multimedia , author =

  77. [85]

    Xu, Mutian and Ding, Runyu and Zhao, Hengshuang and Qi, Xiaojuan , year =. Paconv:

  78. [86]

    Yan, Tianyi and Yin, Junbo and Lang, Xianpeng and Yang, Ruigang and Xu, Cheng-Zhong and Shen, Jianbing , year =

  79. [87]

    Point deformable network with enhanced normal embedding for point cloud analysis , volume =

    Yin, Xingyilang and Yang, Xi and Liu, Liangchen and Wang, Nannan and Gao, Xinbo , year =. Point deformable network with enhanced normal embedding for point cloud analysis , volume =

  80. [88]

    Mm-point:

    Yu, Hai-Tao and Song, Mofei , year =. Mm-point:

  81. [89]

    Point-bert:

    Yu, Xumin and Tang, Lulu and Rao, Yongming and Huang, Tiejun and Zhou, Jie and Lu, Jiwen , year =. Point-bert:

  82. [90]

    2025 , pages =

    AGILE: GIScience Series , author =. 2025 , pages =

    2025

  83. [91]

    Zeng, Ziyin and Dong, Mingyue and Zhou, Jian and Qiu, Huan and Dong, Zhen and Luo, Man and Li, Bijun , year =

  84. [92]

    2022 , keywords =

    Neurocomputing , author =. 2022 , keywords =

    2022

  85. [93]

    Hamdi, Abdullah and Giancola, Silvio and Ghanem, Bernard , year =. Voint

  86. [94]

    Gu, Albert and Dao, Tri , year =. Mamba:

  87. [95]

    Liang, Dingkang and Zhou, Xin and Xu, Wei and Zhu, Xingkui and Zou, Zhikang and Ye, Xiaoqing and Tan, Xiao and Bai, Xiang , year =

  88. [96]

    Rethinking

    Ma, Xu and Qin, Can and You, Haoxuan and Ran, Haoxi and Fu, Yun , year =. Rethinking

  89. [97]

    and Yi, Li and Su, Hao and Guibas, Leonidas J

    Qi, Charles R. and Yi, Li and Su, Hao and Guibas, Leonidas J. , year =

  90. [98]

    Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training , volume =

    Zhang, Renrui and Guo, Ziyu and Gao, Peng and Fang, Rongyao and Zhao, Bin and Wang, Dong and Qiao, Yu and Li, Hongsheng , year =. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training , volume =

  91. [99]

    Pointgpt:

    Chen, Guangyan and Wang, Meiling and Yang, Yi and Yu, Kai and Yuan, Li and Yue, Yufeng , year =. Pointgpt:

  92. [100]

    Pcp-mae:

    Zhang, Xiangdong and Zhang, Shaofeng and Yan, Junchi , year =. Pcp-mae:

  93. [101]

    and Li, Stan Z

    Li, Zhe and Gao, Zhangyang and Tan, Cheng and Ren, Bocheng and Yang, Laurence T. and Li, Stan Z. , year =. General

  94. [102]

    Towards compact 3d representations via point feature enhancement masked autoencoders , volume =

    Zha, Yaohua and Ji, Huizhen and Li, Jinmin and Li, Rongsheng and Dai, Tao and Chen, Bin and Wang, Zhi and Xia, Shu-Tao , year =. Towards compact 3d representations via point feature enhancement masked autoencoders , volume =

  95. [103]

    Su, Kunming and Wu, Qiuxia and Cai, Panpan and Zhu, Xiaogang and Lu, Xuequan and Wang, Zhiyong and Hu, Kun , year =

  96. [104]

    Zhang, Jianhui and Luo, Yizhi and Zhang, Zicheng and Nie, Xuecheng and Li, Bonan , year =

  97. [105]

    Spectral informed mamba for robust point cloud processing , urldate =

    Bahri, Ali and Yazdanpanah, Moslem and Noori, Mehrdad and Dastani, Sahar and Cheraghalikhani, Milad and Hakim, Gustavo Adolfo Vargas and Osowiechi, David and Beizaee, Farzad and Ben Ayed, Ismail and Desrosiers, Christian , year =. Spectral informed mamba for robust point cloud...

  98. [106]

    Wu, Xiaoyang and DeTone, Daniel and Frost, Duncan and Shen, Tianwei and Xie, Chris and Yang, Nan and Engel, Jakob and Newcombe, Richard and Zhao, Hengshuang and Straub, Julian , year =. Sonata:

  99. [107]

    Meta architecture for point cloud analysis , urldate =

    Lin, Haojia and Zheng, Xiawu and Li, Lijiang and Chao, Fei and Wang, Shanshan and Wang, Yan and Tian, Yonghong and Ji, Rongrong , year =. Meta architecture for point cloud analysis , urldate =

  100. [108]

    Pointvector:

    Deng, Xin and Zhang, WenYu and Ding, Qing and Zhang, XinMing , year =. Pointvector:

  101. [109]

    Improved mlp point cloud processing with high-dimensional positional encoding , volume =

    Zou, Yanmei and Yu, Hongshan and Yang, Zhengeng and Li, Zechuan and Akhtar, Naveed , year =. Improved mlp point cloud processing with high-dimensional positional encoding , volume =

  102. [110]

    Wu, Chengzhi and Wan, Yuxin and Fu, Hao and Pfrommer, Julius and Zhong, Zeyun and Zheng, Junwei and Zhang, Jiaming and Beyerer, Jürgen , year =

  103. [111]

    Hengshuang Zhao and Li Jiang and Jiaya Jia and Philip H. S. Torr and Vladlen Koltun , title =. ICCV , pages =

  104. [112]

    Fast point transformer , urldate =

    Park, Chunghyun and Jeong, Yoonwoo and Cho, Minsu and Park, Jaesik , year =. Fast point transformer , urldate =

  105. [113]

    Point transformer v2:

    Wu, Xiaoyang and Lao, Yixing and Jiang, Li and Liu, Xihui and Zhao, Hengshuang , year =. Point transformer v2:

  106. [114]

    2025 , pages =

    CVM , author =. 2025 , pages =

    2025

  107. [115]

    Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding , year=

    Yang, Yu-Qi and Guo, Yu-Xiao and Xiong, Jian-Yu and Liu, Yang and Pan, Hao and Wang, Peng-Shuai and Tong, Xin and Guo, Baining , journal=. Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene Understanding , year=

  108. [116]

    Pointnext:

    Qian, Guocheng and Li, Yuchen and Peng, Houwen and Mai, Jinjie and Hammoud, Hasan and Elhoseiny, Mohamed and Ghanem, Bernard , year =. Pointnext:

  109. [117]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , author =

    Learning view-based graph convolutional network for multi-view. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2022 , pages =

    2022

  110. [118]

    Dosovitskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil , year =. An

  111. [119]

    Liu, Ze and Hu, Han and Cao, Yue and Zhang, Zheng and Tong, Xin , editor =. A. 2020 , pages =

    2020

  112. [120]

    Revisiting

    Uy, Mikaela Angelina and Pham, Quang-Hieu and Hua, Binh-Son and Nguyen, Duc Thanh and Yeung, Sai-Kit , year =. Revisiting

  113. [121]

    SIGGRAPH Asia , author =

    A. SIGGRAPH Asia , author =

  114. [122]

    2025 , journal =

    3DSES: an indoor Lidar point cloud segmentation dataset with real and pseudo-labels from a 3D model , author=. 2025 , journal =

    2025

  115. [123]

    ACM TOMS , author =

    An algorithm for finding best matches in logarithmic expected time , volume =. ACM TOMS , author =. 1977 , note =

    1977

  116. [124]

    3d shapenets:

    Wu, Zhirong and Song, Shuran and Khosla, Aditya and Yu, Fisher and Zhang, Linguang and Tang, Xiaoou and Xiao, Jianxiong , year =. 3d shapenets:

  117. [125]

    2020 , pages =

    Xie, Haozhe and Yao, Hongxun and Zhou, Shangchen and Mao, Jiageng and Zhang, Shengping and Sun, Wenxiu , editor =. 2020 , pages =

    2020

  118. [126]

    ECCV , year=

    Masked Discrimination for Self-Supervised Learning on Point Clouds , author=. ECCV , year=

  119. [127]

    Point cloud pre-training with diffusion models , booktitle =

    Zheng, Xiao and Huang, Xiaoshui and Mei, Guofeng and Hou, Yuenan and Lyu, Zhaoyang and Dai, Bo and Ouyang, Wanli and Gong, Yongshun , year =. Point cloud pre-training with diffusion models , booktitle =

  120. [128]

    Wang, Ziyi and Zhang, Yanran and Zhou, Jie and Lu, Jiwen , year =

  121. [129]

    Generative

    Zhang, Qi and Peng, Jibin and Huang, Zhao and Feng, Wei and Lin, Di , year =. Generative

  122. [130]

    Wang, Song and Liu, Xiaolu and Kong, Lingdong and Xu, Jianyun and Hu, Chunyong and Fang, Gongfan and Li, Wentong and Zhu, Jianke and Wang, Xinchao , year =

  123. [131]

    Oa-cnns:

    Peng, Bohao and Wu, Xiaoyang and Jiang, Li and Chen, Yukang and Zhao, Hengshuang and Tian, Zhuotao and Jia, Jiaya , year =. Oa-cnns:

  124. [132]

    Language-

    Rozenberszki, David and Litany, Or and Dai, Angela , year =. Language-

  125. [133]

    Han, Zongyan and El Amine Boudjoghra, Mohamed and Dong, Jiahua and Wang, Jinhong and Anwer, Rao Muhammad , year =. All in

  126. [134]

    Harnessing

    Chen, Yiyang and Zhao, Shanshan and Duan, Lunhao and Ding, Changxing and Tao, Dacheng , year =. Harnessing

  127. [135]

    Point Cloud Mixture-of-Domain-Experts Model for 3D Self-supervised Learning , booktitle =

    Zha, Yaohua and Dai, Tao and Guo, Hang and Wang, Yanzi and Chen, Bin and Chen, Ke and Xia, Shu-Tao , editor =. Point Cloud Mixture-of-Domain-Experts Model for 3D Self-supervised Learning , booktitle =. 2025 , pages =

    2025

  128. [136]

    ICCV , author =

    Point. ICCV , author =. 2025 , pages =

    2025

  129. [137]

    ICCV , author =

    Towards. ICCV , author =. 2025 , pages =

    2025

  130. [138]

    arXiv preprint arXiv:2308.16532 , year=

    Decoupled local aggregation for point cloud learning , author=. arXiv preprint arXiv:2308.16532 , year=

    arXiv

  131. [139]

    1935 , publisher=

    Hilbert, David , booktitle=. 1935 , publisher=

    1935

  132. [140]

    1966 , publisher=

    A computer oriented geodetic data base and a new technique in file sequencing , author=. 1966 , publisher=

    1966

  133. [141]

    IEEE Transactions on Multimedia , year=

    PMMTalk : Speech-Driven 3D Facial Animation from Complementary Pseudo Multi-modal Features , author=. IEEE Transactions on Multimedia , year=

  134. [142]

    ACM MM , pages=

    PESTalk: Speech-Driven 3D Facial Animation with Personalized Emotional Styles , author=. ACM MM , pages=

  135. [143]

    ICCV , month =

    Zhu, Zhenwei and Yang, Liying and Li, Ning and Jiang, Chaohao and Liang, Yanyan , title =. ICCV , month =. 2023 , pages =

    2023

  136. [144]

    ICCV , month =

    Yang, Liying and Zhu, Zhenwei and Lin, Xuxin and Nong, Jian and Liang, Yanyan , title =. ICCV , month =. 2023 , pages =

    2023

  137. [145]

    Pattern Recognition , volume=

    Garnet: Global-aware multi-view 3d reconstruction network and the cost-performance tradeoff , author=. Pattern Recognition , volume=. 2023 , publisher=

    2023

  138. [146]

    ICCV , month =

    Yang, Liying and Liu, Chen and Zhu, Zhenwei and Liu, Ajian and Ma, Hui and Nong, Jian and Liang, Yanyan , title =. ICCV , month =. 2025 , pages =

    2025

  139. [147]

    arXiv preprint arXiv:2602.18830 , year=

    Spatial-Temporal State Propagation Autoregressive Model for 4D Object Generation , author=. arXiv preprint arXiv:2602.18830 , year=

  140. [148]

    AAAI , pages =

    Xinxing Yu and Ajian Liu and Sunyuan Qiang and Yuzhong Wang and Hui Ma and Yanyan Liang , title =. AAAI , pages =

  141. [149]

    Computational Visual Media , volume=

    FACNet: Feature alignment fast point cloud completion network , author=. Computational Visual Media , volume=. 2025 , publisher=

    2025

  142. [150]

    AAAI , pages=

    MCGS: Markov Chain Gaussian Splatting for Dynamic Scenes Reconstruction , author=. AAAI , pages=

  143. [151]

    NeurIPS , volume=

    Rethinking the compositionality of point clouds through regularization in the hyperbolic space , author=. NeurIPS , volume=

  144. [152]

    Pattern Recognition , volume=

    Hyperbolic prototype rectification for few-shot 3D point cloud classification , author=. Pattern Recognition , volume=. 2025 , publisher=

    2025

  145. [153]

    CoRR , volume =

    Benjy Friedmann and Michael Werman , title =. CoRR , volume =. 2025 , eprinttype =

    2025

  146. [154]

    2026 , journal =

    PointCSP: Cross-Sample Semantic Propagation and Stability Preservation in Self-Supervised Point Cloud Learning , author=. 2026 , journal =

    2026

  147. [155]

    ICMLW , year=

    Point cloud classification with ModelNet40: What is left? , author=. ICMLW , year=

  148. [156]

    ICCV , pages=

    Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data , author=. ICCV , pages=

  149. [157]

    ICML , pages=

    Representation tradeoffs for hyperbolic embeddings , author=. ICML , pages=. 2018 , organization=

    2018

  150. [158]

    CVPR , pages=

    Hyperbolic image segmentation , author=. CVPR , pages=

  151. [159]

    AAAI , pages=

    Hyperbolic hierarchical alignment reasoning network for text-3d retrieval , author=. AAAI , pages=

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.