Pith. sign in

REVIEW 5 major objections 6 minor 16 references

Multi-modal Multi-task Pre-training for Improved Point Cloud Understanding

T0 review · 5 major / 6 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 This paper claims that jointly pre-training a point-cloud encoder on token-level reconstruction, point-level reconstruction, and 2D–3D contrastive learning beats prior self-supervised methods across classification, segmentation…

desk verdict Plausible multi-task recipe with broad, strong-looking completion results, but the joint loss as written contains an undefined label term that makes the self-supervised pre-training unimplementable as described. read the letter →

arxiv 2507.17533 v1 pith:QBSAOHM2 submitted 2025-07-23 cs.CV

Liwen Liu , Weidong Yang , Lipeng Ma , Ben Fei This is my paper
classification cs.CV
keywords pointcloudself-supervisedlearningmulti-modalpre-trainingmulti-taskmaskedautoencodercontrastive3Drepresentationtransformer
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 claims that a point-cloud encoder pre-trained with three self-supervised tasks at once—token-level reconstruction, point-level reconstruction, and multi-modal contrastive learning between 3D shapes and 2D images—transfers better to downstream tasks than encoders pre-trained with any one of those tasks alone. The authors report that MMPT, their multi-task framework, outperforms prior self-supervised methods on object classification (ModelNet40 and ScanObjectNN), part and semantic segmentation, 3D object detection, and shape completion. They also claim that this is the first time multi-task pre-training has been applied to 3D point clouds, and that the whole pipeline runs without any 3D annotations. If true, the framework offers a scalable recipe for turning unlabeled point-cloud and image collections into a general-purpose 3D encoder.

What carries the argument

The central object is MMPT, a multi-modal multi-task pre-training framework whose three pretext tasks share one 3D Transformer encoder. Token-level reconstruction (TLR) follows the masked-autoencoder recipe: random masking of point patches at an 80% ratio, a lightweight PointNet token embedder, a standard Transformer encoder, and a small decoder that predicts masked token coordinates and is trained with Chamfer distance plus a binary focal loss. Point-level reconstruction (PLR) trains a discriminative decoder to distinguish real query points sampled from masked groups from fake queries sampled over the 3D space, and the recovered point cloud is reused as a transformed input for the contrastive branch. Multi-modal contrastive learning (MCL) applies NT-Xent losses both between two transformed views of the same point cloud (intra-modal) and between 3D logits and projected 2D image features (cross-modal). The joint objective is a weighted sum of four loss terms, and the authors' ablations show the three tasks working together outperform any subset.

What would settle it

Examine the training code or supplementary material to find how slabels in Eq. 10 is generated: if it is derived from annotated class labels or from a clustering process that needs ground truth, the self-supervised claim fails. A direct experiment would retrain MMPT with the MoCo term removed and compare the reported gains over Point-MAE; if the gains vanish, the undefined MoCo term is load-bearing, and the method as published is incomplete.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes the following: a shared 3D Transformer encoder, trained jointly on masked token prediction (TLR), masked point-group prediction with real-vs-fake query discrimination (PLR), and intra- plus cross-modal contrastive learning against rendered 2D images (MCL), produces representations that beat the current self-supervised baselines on five downstream tasks. The joint objective combines four loss terms—reconstruction (Chamfer distance plus binary focal loss), a MoCo term, intra-modal NT-Xent, and cross-modal NT-Xent—and the ablations show that all three tasks in combination outperform any one or two of them. The authors further report that a single rendered 2D image per shape suffices for the multi-modal signal, and that adding more views slightly hurts accuracy. The paper positions MMPT as the first multi-task pre-training framework for 3D point clouds, with the trained encoder transferring to classification, part/semantic segmentation, detection, and completion without manual 3D annotation.

Load-bearing premise

The load-bearing premise is that the MoCo loss term in Eq. 10 can be computed self-supervised, but the paper never defines where its labels slabels come from, so if those labels require manual annotation or an unspecified external source, the framework's annotation-free claim cannot be implemented as written.

Editorial extensions

If this is right

  • The pre-trained encoder transfers to classification, part and semantic segmentation, 3D object detection, and shape completion without any 3D annotations, per the paper's five downstream experiments.
  • Multi-task pre-training beats every single-task variant: on ModelNet40 accuracy rises from 93.1% (TLR only) to 93.9%, and on ScanObjectNN from 88.0% to 91.0%.
  • A single rendered 2D view is enough to capture the cross-modal signal; the ablation in Table 12 shows accuracy declines when more than one view is used.
  • The completion gains are the largest: MMPT reports an average CD-ℓ1 of 7.396 on PCN, below SnowflakeNet's 8.362 and PoinTr's 12.211.
  • Because pre-training needs no 3D annotations, the framework can in principle scale to large unlabeled 3D-plus-image datasets.

Reading between the lines

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

  • If the multi-task claim holds, the recipe should generalize to other pretext-task combinations, such as adding point-cloud registration or part-prediction heads, which the paper does not test.
  • The view-number ablation suggests that the averaging of projected features across multiple rendered views may wash out alignment; a natural extension is to replace averaging with a learnable fusion across views.
  • The undefined slabels in the MoCo term leaves the strictly self-supervised claim conditional; testing whether slabels can be generated from the masked reconstruction itself would settle whether the framework is truly annotation-free.
  • The large completion margin relative to the smaller classification margin hints that the PLR task contributes most to generative downstream tasks, a hypothesis the paper's ablations do not separate.
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 / 6 minor

Summary. The paper proposes MMPT, a self-supervised multi-task pre-training framework for 3D point clouds that combines token-level reconstruction (TLR), point-level reconstruction (PLR), and multi-modal contrastive learning (MCL) between point clouds and rendered 2D images. The authors claim this is the first multi-task pre-training framework for 3D point clouds and that it requires no 3D annotations. The trained encoder is transferred to shape classification on ModelNet40 and ScanObjectNN, few-shot classification, part segmentation on ShapeNetPart, scene semantic segmentation on S3DIS, 3D detection on ScanNet V2, and point cloud completion on PCN, MVP, and ShapeNet55/34/21, where it reports state-of-the-art or comparable numbers across these tasks.

Significance. If the claims were fully supported, the paper would make a useful contribution by showing that multiple pre-training objectives can be combined for 3D point cloud representation learning, and the breadth of downstream evaluation is a genuine strength. The supplementary material contains extensive per-category completion results, which is valuable for comparison. However, the current manuscript has unresolved definitional gaps in the central loss function, a direct contradiction between the abstract and Section 3.3 about what PLR does, and an evaluation protocol that selects hyperparameters on the test set. These issues prevent the reader from verifying the central claims, so the significance of the work cannot be assessed from the manuscript as written.

major comments (5)
  1. [Section 3.5, Eq. (10)] The MoCo loss L_MoCo is defined with quantities written as slabels'_k_i and slabels_k_j, and the text states only that 'S3D and Slabels are the 3D logits and labels, respectively.' No construction or source of Slabels is given anywhere in Sections 3.2-3.5, the experiments, or the supplementary material. If Slabels are class labels, the framework violates its promise of operating without any 3D annotations; if they are self-generated, the generation process must be specified precisely. Because L_MoCo is one of the four terms in the joint loss of Eq. (8), the pre-training objective is not fully specified and the method cannot be implemented or reproduced as written.
  2. [Abstract and Section 3.3, Eq. (9)] The abstract states that PLR 'aims to predict the masked point positions directly,' but Section 3.3 says the discriminative decoder 'does not predict the coordinates of masked groups, thus preventing the leakage of position information' and instead distinguishes real queries from fake queries sampled over the 3D space. This is a direct contradiction about what the second pre-training task actually does. In addition, the reconstruction term in Eq. (9) uses Qlabels in Lrec_bce without defining Qlabels; the only label-like object discussed in Section 3.3 is the real/fake distinction, which should be stated explicitly.
  3. [Section 5, Tables 12-13; Section 3.2] The loss-weight ratio (1:1:0.1), the number of rendered views (1), and the masking ratio (gamma=0.8, chosen 'according to the experimental results') are all selected by accuracy on the ModelNet40 test set. Since the same test set is used to report the headline 93.9% accuracy in Table 1, the reported gains are not independent benchmark evaluations. A validation split or a nested selection protocol is needed before the state-of-the-art claims can be taken as reliable.
  4. [Section 3.4, Eqs. (6)-(7) and (11)-(12)] The contrastive losses compute similarities between z3D_i, described as the mean projected vector from the projection head g3D, and s3D_i, described as '3D logits' from the PLR discriminative decoder. The paper never states that these two objects live in the same embedding space, that they have compatible dimensions, or that the logits are normalized before the dot-product similarity is taken. This makes the MCL loss, which is a core component of the method, underspecified.
  5. [Table 1 and Section 4.2.1] Table 1 reports Point-MAE at 93.8 and Point-MAE (Rep.) at 93.1, but the text claims an improvement of 0.8% 'compared to Point-MAE' without acknowledging the 0.7% gap between the published number and the reproduction. This ambiguity matters because the headline gain is 0.8% over the reproduced baseline but only 0.1% over the published baseline, and no explanation is given for the reproduction gap.
minor comments (6)
  1. [References and related work] Point-MAE is cited inconsistently as [Liu et al., 2022] in Sections 1 and 2.1 and as [Pang et al., 2022] in Sections 3.2 and 4.2.1, while [Liu et al., 2022] in the reference list is actually MaskPoint; this needs to be corrected throughout.
  2. [Figure 1] The framework figure includes labels 'Loss1' and 'Loss2' that are not defined in the caption or in the surrounding text.
  3. [Section 3.4] The sentence 'As illustrated in Fig 3' appears in the description of MCL, but the MCL pipeline is shown in Figure 4, not Figure 3.
  4. [Table 9] The header of Table 9 is malformed: the column groups for F1-Avg, CD-S, CD-M, CD-H, and CD-Avg are not visually separated, making the table very difficult to read.
  5. [Section 5] The phrase 'multi-model and multi-task' should be 'multi-modal and multi-task' to match the terminology used in the rest of the paper.
  6. [Section 4.2.7] The text references 'Figure 7' for both the PCN and MVP qualitative results; the second reference should point to a distinct figure, since Figure 7 appears to show only one dataset.

Circularity Check

0 steps flagged · score 2.0 of 10

No constructional circularity: the reported gains are empirical transfer results from a defined multi-task loss (underspecified at Eq. 10); mild benchmark-tuning and the undefined Slabels are reproducibility risks, not circular derivation.

full rationale

The central derivation chain is empirical rather than constructional. The pre-training objective in Eq. (8) is a weighted sum of four terms, each specified (apart from the undefined Slabels issue noted below) in Eqs. (9)-(12): reconstruction of masked patches, a MoCo-style contrastive term on the PLR logits, intra-modal NT-Xent, and cross-modal NT-Xent. The encoder is then fine-tuned and evaluated on external splits of ModelNet40, ScanObjectNN, ShapeNetPart, S3DIS, ScanNet V2, and the completion benchmarks. There is no equation in which a predicted quantity is defined in terms of the very quantity it predicts: TLR predicts masked patches from visible patches, PLR discriminates real versus fake queries, and MCL aligns augmented 3D views with rendered 2D images. The self-citations (Fei et al. 2022-2025, Liu et al. 2025) appear in the related-work and method-inspiration context and are not load-bearing; no uniqueness theorem or fitted value from those papers is imported to force the present result. Two concerns are flagged, but they are not constructional circularity. First, hyperparameters (loss ratio 1:1:0.1 in Table 13; one rendered view in Table 12; mask ratio 0.8 in Section 3.2) are selected using ModelNet40 accuracy, and the same ModelNet40 accuracy is then reported as a headline result; this is selection-on-test rather than independent validation. Second, Eq. (10) introduces Slabels with only the note 'where S3D and Slabels are the 3D logits and labels, respectively'; no source or construction of these labels is given, leaving the self-supervised objective underspecified and unverifiable. These are correctness and reproducibility defects that raise the burden slightly above zero but do not make the central transfer claim equivalent to its inputs.

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

The method is built from existing components: Point-BERT token prediction, Point-MAE-style reconstruction, and CrossPoint-style contrastive learning. The only new element is the specific loss combination, and the most fragile part is the undefined label source in the MoCo term. The free parameters are tuned against test benchmarks, which weakens the reported improvements.

free parameters (4)
  • Masking ratio gamma = 0.8
    Set to 0.8 according to experimental results in Section 3.2; the value is chosen empirically rather than derived.
  • Loss weight ratio alpha:beta:gamma = 1:1:0.1
    Selected by maximizing ModelNet40 test accuracy in Table 13, a test-set based selection that is reported as the final configuration.
  • Contrastive temperature tau = not specified
    Appears in Eqs. 6 and 7; no value or selection procedure is reported, so it remains a free choice.
  • Number of rendered views = 1
    Ablation in Table 12 shows one view is best on ModelNet40; the final model uses one view, tuned on the evaluation set.
assumptions (4)
  • domain assumption Token-level reconstruction via cross-entropy is a useful pre-training signal for point cloud transformers
    Stated in Section 3.2 as a commonly employed method, inherited from Point-BERT without independent justification in this paper.
  • domain assumption Training the decoder to distinguish real and fake query points improves geometric understanding
    Section 3.3 assumes this discriminative task conveys shape information, but no analysis or evidence is provided beyond the final benchmark numbers.
  • domain assumption Contrastive alignment of 2D image features and 3D point features transfers useful semantics
    Section 3.4 relies on established contrastive learning assumptions, e.g., SimCLR and CrossPoint, without new evidence.
  • ad hoc to paper Labels are available for the MoCo term in a self-supervised framework
    Equation 10 uses slabels without defining them; since the paper claims no annotations are used, this is an unsupported premise specific to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-modal Multi-task Pre-training for Improved Point Cloud Understanding." pith.science (2026). https://pith.science/paper/QBSAOHM2

@misc{pith2026250717533,
  author       = {Pith},
  title        = {Pith review of: Multi-modal Multi-task Pre-training for Improved Point Cloud Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QBSAOHM2}},
  note         = {Machine review of arXiv:2507.17533}
}
read the original abstract

Recent advances in multi-modal pre-training methods have shown promising effectiveness in learning 3D representations by aligning multi-modal features between 3D shapes and their corresponding 2D counterparts. However, existing multi-modal pre-training frameworks primarily rely on a single pre-training task to gather multi-modal data in 3D applications. This limitation prevents the models from obtaining the abundant information provided by other relevant tasks, which can hinder their performance in downstream tasks, particularly in complex and diverse domains. In order to tackle this issue, we propose MMPT, a Multi-modal Multi-task Pre-training framework designed to enhance point cloud understanding. Specifically, three pre-training tasks are devised: (i) Token-level reconstruction (TLR) aims to recover masked point tokens, endowing the model with representative learning abilities. (ii) Point-level reconstruction (PLR) is integrated to predict the masked point positions directly, and the reconstructed point cloud can be considered as a transformed point cloud used in the subsequent task. (iii) Multi-modal contrastive learning (MCL) combines feature correspondences within and across modalities, thus assembling a rich learning signal from both 3D point cloud and 2D image modalities in a self-supervised manner. Moreover, this framework operates without requiring any 3D annotations, making it scalable for use with large datasets. The trained encoder can be effectively transferred to various downstream tasks. To demonstrate its effectiveness, we evaluated its performance compared to state-of-the-art methods in various discriminant and generative applications under widely-used benchmarks.

Figures

Figures reproduced from arXiv: 2507.17533 by the authors.

Figure 1
Figure 1. The overall framework of our MMPT. Our MMPT is a novel multi-task pre-training framework that consists [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Illustrations of the masked point tokens prediction task, which enhances the classification capabilities of the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustrations of the masked point groups prediction task, which enhances the generation ability of backbone. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The pipeline of MCL, which can improve the classification ability of the network. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization comparison of semantic segmentation on ShapeNetPart dataset by different methods. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visualization comparison of point cloud completion on PCN dataset by different methods. From left to right: [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Visualization of qualitative comparison results on the MVP dataset by different methods. From top to bottom: [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visualization of qualitative comparison results on the ShapeNet55/21 dataset by different methods. From [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: t-SNE visualization of the features learned from ModelNet40 and ScanObjectNN (OBJ-ONLY). [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

16 extracted references · 10 canonical work pages

  1. [1]

    Comprehensive review of deep learning-based 3d point cloud completion processing and analysis

    Ben Fei, Weidong Yang, Wen-Ming Chen, Zhijun Li, Yikang Li, Tao Ma, Xing Hu, and Lipeng Ma. Comprehensive review of deep learning-based 3d point cloud completion processing and analysis. IEEE Transactions on Intelligent Transportation Systems, 2022a. Ben Fei, Weidong Yang, Liwen Liu, Tianyue Luo, Rui Zhang, Yixuan Li, and Ying He. Self-supervised learning...

    arXiv

  2. [4]

    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, 2022b. Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2...

    arXiv

  3. [7]

    Image2point: 3d point-cloud understanding with 2d image pretrained models

    Chenfeng Xu, Shijia Yang, Tomer Galanti, Bichen Wu, Xiangyu Yue, Bohan Zhai, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Image2point: 3d point-cloud understanding with 2d image pretrained models. arXiv preprint arXiv:2106.04180,

    arXiv

  4. [8]

    P2P: Tuning Pre-trained Image Models for Point Cloud Analysis with Point-to-Pixel Prompting

    Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. P2p: Tuning pre-trained image models for point cloud analysis with point-to-pixel prompting. arXiv preprint arXiv:2208.02812,

    work page Pith review arXiv

  5. [9]

    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. In Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part II, pages 604–621. Springer,

    work page 2022

  6. [10]

    Pvt: Point-voxel transformer for 3d deep learning

    Cheng Zhang, Haocheng Wan, Shengqiang Liu, Xinyi Shen, and Zizhao Wu. Pvt: Point-voxel transformer for 3d deep learning. arXiv preprint arXiv:2108.06076, 2,

    arXiv

  7. [11]

    Asfm-net: Asymmetrical siamese feature matching network for point completion

    Yaqi Xia, Yan Xia, Wei Li, Rui Song, Kailang Cao, and Uwe Stilla. Asfm-net: Asymmetrical siamese feature matching network for point completion. In Proceedings of the 29th ACM International Conference on Multimedia , pages 1938–1947,

    work page 1938

  8. [16]

    We fine-tune our model on the point cloud completion benchmarks for 200 epochs

    For point cloud completion task, we utilize a standard Transformer encoder and a powerful Transformer-based decoder devised in SnowflakeNet [Xiang et al., 2021]. We fine-tune our model on the point cloud completion benchmarks for 200 epochs. 21 arXiv Template A PREPRINT Table 14: Point cloud completion on PCN in terms of F-score@1% (higher is better). F-s...

    work page arXiv 2021

Show all 16 references
  1. [23]

    We follow previous works and use 1024 points with coordinate information as the input [Yu et al., 2022, Lu et al., 2022, Gao et al., 2022]

    A Datasets ModelNet40 [Wu et al., 2015] dataset comprises 12,311 CAD models from 40 object categories, with 9,843 samples used for training and 2,468 samples for testing. We follow previous works and use 1024 points with coordinate information as the input [Yu et al., 2022, Lu...

    2015

  2. [32]

    Similar to PointNet [Qi et al., 2017a], we sample 2,048 points from each model

    For the fine-grained 3D recognition task of part segmentation, we use ShapeNetPart [Yi et al., 2016], which comprises 16,881 objects of 2,048 points from 16 categories with 50 parts in total. Similar to PointNet [Qi et al., 2017a], we sample 2,048 points from each model. For P...

    2016

  3. [2008]

    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,

    1912

  4. [2018]

    Grnet: Gridding residual network for dense point cloud completion

    Haozhe Xie, Hongxun Yao, Shangchen Zhou, Jiageng Mao, Shengping Zhang, and Wenxiu Sun. Grnet: Gridding residual network for dense point cloud completion. In European Conference on Computer Vision, pages 365–381. Springer, 2020b. Wentao Yuan, Tejas Khot, David Held, Christoph M...

    2018

  5. [2019]

    Pointcontrast: Unsupervised pre-training for 3d point cloud understanding

    Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 5...

    2020

  6. [2020]

    Gs-pt: Exploiting 3d gaussian splatting for comprehensive point cloud understanding via self-supervised learning

    Keyi Liu, Yeqi Luo, Weidong Yang, Jingyi Xu, Zhijun Li, Wen-Ming Chen, and Ben Fei. Gs-pt: Exploiting 3d gaussian splatting for comprehensive point cloud understanding via self-supervised learning. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Sign...

    2025 arXiv

  7. [2021]

    Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining

    Yueh-Cheng Liu, Yu-Kai Huang, Hung-Yueh Chiang, Hung-Ting Su, Zhe-Yu Liu, Chin-Tang Chen, Ching-Yu Tseng, and Winston H Hsu. Learning from 2d: Contrastive pixel-to-point knowledge transfer for 3d pretraining. arXiv preprint arXiv:2104.04687,

    arXiv

  8. [2022]

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

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. arXiv preprint arXiv:2205.14401, 2022a. Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof...

    arXiv

Pith tools

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