Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Exploring Semantic Masked Autoencoder for Self-supervised Point Cloud Understanding

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Whole-part masking lifts point cloud accuracy past 90 percent.

desk verdict Plausible and useful incremental extension of point-cloud MAE; the semantic-masking attribution needs a control that randomizes prototype identity while matching mask distribution. read the letter →

arxiv 2506.21957 v1 pith:OZBHPKDT submitted 2025-06-27 cs.CV

Yixin Zha , Chuxin Wang , Wenfei Yang , Tianzhu Zhang This is my paper
classification cs.CV
keywords pointcloudself-supervisedlearningmaskedautoencodercomponentsemanticslearnableprototypesmodelingprompttuning
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

The paper targets a blind spot in masked point modeling: random masking leaves part of a component visible, so the network can reconstruct a masked patch from adjacent surface structure without ever reasoning about the component as a whole. Its claim is that semantic masking, in which whole components such as airplane wings are hidden at once, forces the model to infer the missing part from the rest of the object and thereby learn component semantics without labels. To get the masking, learnable prototypes group point tokens into components, and the same prototypes are reused as prompts during fine-tuning. On the real-world ScanObjectNN OBJ-ONLY split, the method lifts PointMAE from 88.29 to 90.30 percent accuracy and PointMamba from 88.47 to 90.59 percent, with consistent gains on ModelNet40, ShapeNetPart, and few-shot classification.

What carries the argument

The load-bearing object is a bank of $Q$ learnable prototypes $p \in \mathbb{R}^{Q\times C}$ together with the similarity map $S_m = \mathrm{Softmax}(\hat{T}^e(\hat{p})^T/\sqrt{d})$ that assigns each point token to the prototype it resembles most. Each prototype is meant to become one semantic component, such as a wing or a fuselage; the prototypes are updated by cross-attention over encoder tokens, kept distinct by a contrastive loss, and grounded by reconstructing the original points from each group's prototype plus position embedding. A non-parametric $k$-norm expansion widens the encoder's receptive field so that dispersed structures like wings can be gathered into one prototype. This machinery does double duty: it defines the semantic groups used to mask complete components during pre-training, and it supplies the prompts injected during fine-tuning.

What would settle it

Retrain the full pipeline with the similarity-based grouping replaced by a random-but-fixed partition of the same tokens into Q groups, keeping every other module and loss identical; if ScanObjectNN OBJ-ONLY accuracy stays near 90.30 rather than falling back toward the 88.29 baseline, then semantic grouping is not what causes the gain.

Watch

Extended reading notes

Core claim

The discovery is that the representational weakness of masked point autoencoders comes from the masking schedule, not only from the reconstruction objective, and that fixing the schedule with semantic priors fixes the representations. The paper shows that when a set of learnable prototypes is trained, through cross-attention and prototype-guided point reconstruction, to capture component-level semantics, and when masking then hides one or more complete components instead of random blocks, the pre-trained encoder acquires features that correlate with object parts rather than merely with positions. The learned prototypes also serve as semantic prompts at fine-tuning time, improving downstream tasks. The paper reports this pattern across two backbones, a Transformer-based PointMAE and a Mamba-based PointMamba, on object classification, part segmentation, and few-shot classification.

Load-bearing premise

The entire gain depends on the learnable prototypes settling into semantic components that are consistent across object categories; if the grouping is actually driven by spatial position or local patch geometry, then 'component masking' is just block masking with extra parameters and the claimed reason for improvement collapses.

Editorial extensions

If this is right

  • Pre-trained point cloud models can improve by replacing random masking with prototype-driven component masking, without any additional supervision or cross-modal data.
  • The learned prototypes are reusable: the same vectors that define the masking groups can be plugged into downstream heads as semantic prompts and produce further gains.
  • The gains hold for both Transformer and Mamba backbones, suggesting semantic component masking is a general plug-in for masked point modeling rather than a fix specific to one architecture.
  • Few-shot classification improves along with full fine-tuning, which is consistent with the model having learned transferable part semantics during pre-training.
  • Part segmentation on ShapeNetPart improves by roughly 0.4 to 0.5 mIoU, showing the semantic component prior helps dense prediction as well as global classification.

Reading between the lines

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

  • An extension the paper leaves implicit is an adaptive masking policy that selects which components to hide based on the object's own grouping, rather than sampling a fixed number of groups; the reported ablations suggest masking strategy matters as much as which groups are chosen.
  • The k-norm expansion is described only as a supplementary detail; a direct ablation of it would test whether dispersed-structure grouping, and thus the semantic story, is what carries the gain rather than the extra parameters.
  • Because prototypes are shared across all objects in the pretraining set, one can test whether the learned components are category-general parts or category-specific clusters; this distinction would decide how far the semantic prior transfers to unseen categories.
  • The same prompt mechanism could be compared against a prompt branch trained without prototypes during fine-tuning, which would isolate whether the semantic content of the prompts or merely their extra parameters drives the downstream gain.
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

4 major / 4 minor

Summary. The paper proposes Semantic Masked Autoencoder (SMAE) for self-supervised point cloud pre-training. It introduces a Prototype-based Component Semantic Modeling (PCSM) module that learns a set of Q prototypes by cross-attending to complete point cloud tokens, reconstructs the point cloud group-wise through prototype-conditioned heads, and computes a similarity map that assigns each token to a prototype. This map drives a Component Semantic-enhanced Masking (CSeM) strategy that masks complete semantic components instead of random blocks, and the same prototypes are reused by a Component Semantic-enhanced Prompt-tuning (CSeP) strategy at fine-tuning time. The method is applied on top of PointMAE and PointMamba, pre-trained on ShapeNet, and evaluated by fine-tuning on ScanObjectNN, ModelNet40, ShapeNetPart, and ModelNet40 few-shot classification. The main reported gains are 2.01 and 2.12 percentage points on ScanObjectNN OBJ-ONLY for the two backbones, with smaller gains on other tasks.

Significance. If the central mechanism holds, this is a useful step for masked point modeling: it directly targets the known limitation that random masking leaves local structure intact and does not force reasoning about complete object components, and it demonstrates positive transfer on two different backbones across four downstream tasks. The evaluation is external transfer accuracy, so there is no circularity with the target metrics, and the paper includes ablations of the PCSM, masking, and prompt-tuning components. The few-shot experiments in Table 3 report mean and standard deviation over ten runs, which is a strength. However, the manuscript as submitted does not yet establish that the gains come from the semantic identity of the grouping rather than from mask geometry or the auxiliary losses, and the missing k-norm definition, omitted hyperparameters, and inconsistent baseline numbers make the mechanism and the ablations difficult to evaluate. No code or supplementary material is provided to close these gaps.

major comments (4)
  1. [Section 3.2, Eqs. (2)-(5) and Tables 4-5] The central claim is that masking complete semantic components, rather than random blocks, improves representations, but the ablation design does not isolate the semantic identity of the grouping. In Table 5, the RandM/RandBM rows and the CSeM row differ in mask geometry, and the text does not state whether the RandM and RandBM rows include the PCSM auxiliary losses (PPR and contrastive). Even under the charitable reading that all rows use the same PCSM module, the component size distribution and spatial extent of the masks are not matched across conditions. The reported gain could therefore come from masking larger contiguous regions, from the extra prototype reconstruction and contrastive losses, or from the semantic grouping itself. A control that assigns tokens to Q groups randomly while preserving the group size distribution and using the same mask ratios and the same PCSM losses is required, together with an ablation that applies component-level masks to a non-semantic geometric grouping, before the central attribution claim is established.
  2. [Section 3.2, Eq. (5)] Equation (5) as written is degenerate. With d(·) described as a "distance measurement", the numerator exp(d(\ hat p_i, \ hat p_i)/ε) is identically exp(0)=1, so the loss reduces to a repulsion term among prototypes and no positive-pair signal is defined; if a similarity such as cosine similarity is intended instead, the text must say so and specify what constitutes a positive pair. Since Table 4 shows that the contrastive term participates in the best configuration (the full row reaches 89.51/93.51 versus 87.89/92.69 for the row without prototype reconstruction), this equation needs to be corrected or precisely specified before the PCSM module can be properly evaluated.
  3. [Section 3.2, k-norm expansion] The "non-parametric k-norm process" is load-bearing for the paper's explanation that prototypes can capture dispersed structures such as airplane wings, but the manuscript says only that "the details of k-norm are presented in supplementary materials" and gives no definition, no depth H, and no analysis in the main text. Without this definition, the reader cannot check whether the learned grouping is semantically consistent across object categories or whether it is largely driven by position embeddings, in which case CSeM would degenerate to block masking with extra parameters. The main text should include the k-norm formulation, the value of H used in the experiments, and a sensitivity analysis of grouping quality with respect to H.
  4. [Section 4.2, Tables 1-6] The experimental reporting is not sufficient to support the magnitude of the claimed gains. The vanilla PointMAE OBJ-ONLY accuracy is reported as 88.29 in Table 1, 88.76 in Table 4 (first row), 88.79 or 88.28 in Table 5, and 88.98 in Table 6 (first row), with no explanation for the variation; moreover, Table 6's no-CSeP row (88.98) does not match the CSeM result of 89.51 in Table 1, so the CSeP ablation does not start from the CSeM condition. Tables 1 and 2 report single-run accuracies with no error bars, and the pre-training hyperparameters are not given (Q, ε, component mask ratio, per-component random mask ratio, loss weights, and k-norm depth H are all absent). Because the main claimed improvements are only 0.4-2.1 percentage points, multiple seeds with mean and standard deviation, a consistent baseline across all ablation tables, and a full pre-training configuration (or released code) are needed to assess whether the improvements are significant and reproducible.
minor comments (4)
  1. [Section 4.2, part segmentation paragraph] The sentence "The 'PointMAE + CLIP' combination outperforms all methods in various experimental settings" refers to a model that is not used anywhere in this paper and appears to be a leftover from another manuscript; it should be removed or replaced with the actual ShapeNetPart results reported in Table 2.
  2. [Figure 5] The caption "Raw Points Component Points by LSaM" uses the undefined acronym LSaM, while the surrounding text attributes the components to the proposed method; the caption should be corrected to match the method name or the acronym should be defined.
  3. [Section 4.2 and throughout] There are numerous grammatical and typographical errors that should be fixed, including "default data argumentation" (should be "default data augmentation"), "boost the performence", "the proposed paradigm boost the performence of 3D pre-trained on the dataset", and "All above methods are rely on global random masking strategy" in Section 2.2.
  4. [Sections 2.2, 3.2, and 4.1] Notation is inconsistent across the paper ("Point-MAE" vs "PointMAE", "CrossNet" vs "CrossNets"), and the text repeatedly defers to "supplementary materials" for k-norm details, architecture details, and point cloud details, but no supplementary file is provided with the manuscript; these references should be resolved in the revision.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the reported gains are external transfer accuracies and the prototypes are not defined in terms of the downstream metrics.

full rationale

The paper's derivation chain is self-contained in the sense required by this pass. The central claimed result is that masking complete semantic components, selected by learned prototypes, improves self-supervised point cloud representations, with downstream accuracy on ScanObjectNN, ModelNet40, and ShapeNetPart used as evidence. None of the paper's equations define the evaluated metrics in terms of the prototypes, and no fitted parameter is relabeled as a prediction: the prototypes are trained with reconstruction and contrastive losses (Eqs. 2-5), the masking is generated from the learned similarity map (Eq. 3), and the improvement is measured on held-out classification and segmentation tasks. The paper does cite prior work, including Point-NN for the non-parametric k-norm idea, but this citation is not load-bearing for the core claim and no uniqueness theorem is imported from the authors' own prior work. The main scientific weakness is an attribution confound: the CSeM ablation (Table 5) does not control for mask size, component shape, or the extra PCSM losses, so the gain ascribed to 'semantic' masking could partly come from masking larger contiguous regions or from the additional losses. However, a confound is a correctness risk, not a circularity, because the accuracy numbers are external transfer results that could in principle have gone the other way. There is no step in which a result is equivalent to its input by construction, and no self-citation chain that forces the conclusion. Score 0 with no circular steps.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on prototype-based component grouping; the main free parameters (Q, epsilon, mask ratios, loss weights, k-norm depth) are not reported, and the semantic-consistency assumption is not independently validated.

free parameters (5)
  • Number of prototypes Q
    Prototype set p in R^{QxC} (Section 3.2); Q controls component granularity and is not reported in main text.
  • Contrastive temperature epsilon
    Used in L_cont in Eq. 5; not specified.
  • Component masking ratio and per-component random mask ratio
    CSeM randomly selects token groups to mask and applies random masking on remaining groups (Section 3.3); the ratios are not reported.
  • Loss weights for L_3D, L_proto, L_cont
    The total pre-training loss is not written down; balancing coefficients are unstated.
  • k-norm expansion depth H
    Figure 2 shows k-norm repeated xH, but H is not given; Section 3.2 defers to supplementary.
assumptions (4)
  • domain assumption FPS and k-NN tokenization preserves object components and supports semantic grouping.
    Token embedding in Section 3.1 builds patches from FPS centers and k-NN; the grouping in CSeM assumes these patches align with semantic parts.
  • domain assumption Q learnable prototypes converge to semantically consistent components through L_proto and L_cont.
    Section 3.2; this is the paper's core modeling assumption and is not independently verified beyond qualitative Figure 5.
  • ad hoc to paper k-norm expansion gives prototypes access to distant structures and is necessary for dispersed components.
    Section 3.2 introduces k-norm 'inspired by Point-NN' with details only in absent supplementary; the claim that it fixes dispersed-structure modeling is unverified.
  • standard math Chamfer distance and standard cross-attention are valid reconstruction and aggregation tools.
    Eq. 1 and Eq. 2 inherit from prior MAE methods; no issue.
invented entities (1)
  • Learnable component-semantic prototypes p and updated p-hat
    purpose: Capture local component semantics, drive semantic masking, and serve as prompts at fine-tuning
    Prototypes are trained on ShapeNet without supervision and visualized qualitatively in Figure 5; there is no external benchmark showing they correspond to ground-truth parts, so they are not independently falsifiable outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Semantic Masked Autoencoder for Self-supervised Point Cloud Understanding." pith.science (2026). https://pith.science/paper/OZBHPKDT

@misc{pith2026250621957,
  author       = {Pith},
  title        = {Pith review of: Exploring Semantic Masked Autoencoder for Self-supervised Point Cloud Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZBHPKDT}},
  note         = {Machine review of arXiv:2506.21957}
}
read the original abstract

Point cloud understanding aims to acquire robust and general feature representations from unlabeled data. Masked point modeling-based methods have recently shown significant performance across various downstream tasks. These pre-training methods rely on random masking strategies to establish the perception of point clouds by restoring corrupted point cloud inputs, which leads to the failure of capturing reasonable semantic relationships by the self-supervised models. To address this issue, we propose Semantic Masked Autoencoder, which comprises two main components: a prototype-based component semantic modeling module and a component semantic-enhanced masking strategy. Specifically, in the component semantic modeling module, we design a component semantic guidance mechanism to direct a set of learnable prototypes in capturing the semantics of different components from objects. Leveraging these prototypes, we develop a component semantic-enhanced masking strategy that addresses the limitations of random masking in effectively covering complete component structures. Furthermore, we introduce a component semantic-enhanced prompt-tuning strategy, which further leverages these prototypes to improve the performance of pre-trained models in downstream tasks. Extensive experiments conducted on datasets such as ScanObjectNN, ModelNet40, and ShapeNetPart demonstrate the effectiveness of our proposed modules.

Figures

Figures reproduced from arXiv: 2506.21957 by the authors.

Figure 1
Figure 1. (a) Features distribution. We visualize the features be￾fore and after fine-tune. The feature colors are transformed into fea￾ture space using PCA, where the same color indicates feature consis￾tency. (b) Different mask and corresponding reconstruction pro￾cesses. Random masking masks only local blocks, while semantic masking masks complete components based on semantics (wings). Points in gray circles are masked reg… view at source ↗
Figure 2
Figure 2. Pipeline of our proposed framework. The complete tokens T are input into the encoder to acquire T e , which are fed into PCSM to generate local semantic-enhanced prototypes pˆ and enhanced tokens Tˆe. The PPR guides the learning of pˆ to acquire informative prototypes, and Sm are solved before Semantic Grouping. With Sm, we can utilize CSeM to generate semantic-correlated masks. The generated masks are capable of co… view at source ↗
Figure 3
Figure 3. Details of CSeM. We first segment point clouds into sev￾eral components according to local semantics and random select a few of them as masked components. Then we adopt random mask￾ing on each remaining component separately [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Details of CSeP. (a) The figure illustrates a structure widely used for downstream classification tasks. (b) In the figure, we introduce the PCSM module during the finetuning stage to gen￾erate semantic-enhanced prototypes. 3.3 Component Semantic-enhanced Masking Strat…
Figure 5
Figure 5. Figure 5: Visualization of Components. We visualized the com￾ponents obtained through semantic grouping, which clearly show strong semantic relevance. Part Segmentation on ShapeNetPart Dataset. We conduct part segmentation experiments on the challenging ShapeNetPart [Yi et al., …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. StruMamba3D: Exploring Structural Mamba for Self-supervised Point Cloud Representation Learning

    cs.CV 2025-06 conditional novelty 7.0 of 10

    A self-supervised point cloud model that encodes spatial structure into SSM latent states and adapts state-update scale to input length achieves new SOTA on ScanObjectNN and ModelNet40.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

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

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

    work page 2022

  2. [7]

    An image is worth 16x16 words: Transformers for image recognition at scale

    [Dosovitskiy et al., 2020] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Min- derer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929,

    arXiv 2020

  3. [8]

    A point set generation network for 3d object re- construction from a single image

    [Fan et al., 2017] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object re- construction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 605–613,

    work page 2017

  4. [10]

    Masked au- toencoders are scalable vision learners

    [He et al., 2022] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked au- toencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009,

    work page 2022

  5. [11]

    Pointmamba: A simple state space model for point cloud analysis

    [Liang et al., 2024] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaoqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739,

    arXiv 2024

  6. [14]

    [Pang et al., 2022] Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. arXiv e-prints,

    work page 2022

  7. [15]

    Con- trast with reconstruct: Contrastive 3d representation learn- ing guided by generative pretraining

    [Qi et al., 2023] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Con- trast with reconstruct: Contrastive 3d representation learn- ing guided by generative pretraining. In International Conference on Machine Learning , pages 28223–28243. PMLR,

    work page 2023

  8. [16]

    Improving language understanding by generative pre-training,

    [Radford et al., 2018] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training,

    work page 2018

Show all 31 references
  1. [17]

    Language models are unsupervised multitask learners

    [Radford et al., 2019] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9,

    2019

  2. [18]

    Geomae: Masked geometric target prediction for self-supervised point cloud pre-training

    [Tian et al., 2023] Xiaoyu Tian, Haoxi Ran, Yue Wang, and Hang Zhao. Geomae: Masked geometric target prediction for self-supervised point cloud pre-training. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13570–13580,

    2023

  3. [19]

    Revis- iting point cloud classification: A new benchmark dataset and classification model on real-world data

    [Uy et al., 2019] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revis- iting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceed- ings of the IEEE/CVF international conference on com...

    2019

  4. [20]

    Extract- ing and composing robust features with denoising autoen- coders

    [Vincent et al., 2008] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extract- ing and composing robust features with denoising autoen- coders. In Proceedings of the 25th international confer- ence on Machine learning , pages 1096–1103,

    2008

  5. [22]

    3d shapenets: A deep representation for volumetric shapes

    [Wu et al., 2015] 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 com- puter vision and pattern recognition , pages 1912–1920,

    2015

  6. [23]

    Self-supervised intra-modal and cross- modal contrastive learning for point cloud understanding

    [Wu et al., 2023] Yue Wu, Jiaming Liu, Maoguo Gong, Peiran Gong, Xiaolong Fan, A Kai Qin, Qiguang Miao, and Wenping Ma. Self-supervised intra-modal and cross- modal contrastive learning for point cloud understanding. IEEE Transactions on Multimedia, 26:1626–1638,

    2023

  7. [24]

    Point- contrast: Unsupervised pre-training for 3d point cloud un- derstanding

    [Xie et al., 2020] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Point- contrast: Unsupervised pre-training for 3d point cloud un- derstanding. In Computer Vision–ECCV 2020: 16th Eu- ropean Conference, Glasgow, UK, August 23–28, 2020, Proceedi...

    2020

  8. [25]

    Gd-mae: generative decoder for mae pre-training on lidar point clouds

    [Yang et al., 2023] Honghui Yang, Tong He, Jiaheng Liu, Hua Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wanli Ouyang. Gd-mae: generative decoder for mae pre-training on lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pa...

    2023

  9. [26]

    A scalable active framework for region annotation in 3d shape collec- tions

    [Yi et al., 2016] 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 collec- tions. ACM Transactions on Graphics (ToG), 35(6):1–12,

    2016

  10. [28]

    Self-supervised pretraining of 3d features on any point-cloud

    [Zhang et al., 2021] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 10252–10263,

    2021

  11. [30]

    Parameter is not all you need: Starting from non- parametric networks for 3d point cloud analysis

    [Zhang et al., 2023a] Renrui Zhang, Liuhui Wang, Ziyu Guo, Yali Wang, Peng Gao, Hongsheng Li, and Jianbo Shi. Parameter is not all you need: Starting from non- parametric networks for 3d point cloud analysis. arXiv preprint arXiv:2303.08134,

    arXiv

  12. [31]

    Point transformer

    [Zhao et al., 2021] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259–16268, 2021

    2021

  13. [2008]

    Dynamic graph cnn for learning on point clouds

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

    2019

  14. [2015]

    Pimae: Point cloud and image inter- active masked autoencoders for 3d object detection

    [Chen et al., 2023] Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image inter- active masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patter...

    2023

  15. [2016]

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

    [Yu et al., 2022] 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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313– 19322,

    2022

  16. [2017]

    Mamba: Linear- time sequence modeling with selective state spaces

    [Gu and Dao, 2023] Albert Gu and Tri Dao. Mamba: Linear- time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752,

    2023 arXiv

  17. [2018]

    [Dong et al., 2023] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pre- trained 2d image transformers help 3d representation learning? In The Eleventh International Conference on Lear...

    2023

  18. [2019]

    Masked discrimination for self-supervised learning on point clouds

    [Liu et al., 2022] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vi- sion, pages 657–675. Springer,

    2022

  19. [2020]

    Shapenet: An information-rich 3d model repository

    [Chang et al., 2015] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012,

    2015 arXiv

  20. [2021]

    Point-m2ae: multi-scale masked autoen- coders for hierarchical point cloud pre-training

    [Zhang et al., 2022] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoen- coders for hierarchical point cloud pre-training. Advances in neural information processing systems , 35:27061– 27074,

    2022

  21. [2022]

    Language models are few-shot learners

    [Brown et al., 2020] Tom Brown, Benjamin Mann, Nick Ry- der, Melanie Subbiah, Jared D Kaplan, Prafulla Dhari- wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing sys- tems,...

    2020

  22. [2023]

    Bert: Pre-training of deep bidirectional transformers for language understand- ing

    [Devlin et al., 2018] Jacob Devlin, Ming-Wei Chang, Ken- ton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understand- ing. arXiv preprint arXiv:1810.04805,

    2018 arXiv

  23. [2024]

    Relation-shape convolutional neu- ral network for point cloud analysis

    [Liu et al., 2019] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neu- ral network for point cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8895–8904,

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.