Pith. sign in

REVIEW 4 major objections 6 minor 75 references

A Cross Branch Fusion-Based Contrastive Learning Framework for Point Cloud Self-supervised Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Contrastive learning for point clouds can fuse information between branches before the loss end, and PoCCA claims this yields better self-supervised representations than prior single-modal methods.

desk verdict Useful cross-branch fusion idea for point-cloud SSL, but the SOTA claim outruns the paper's own tables. read the letter →

arxiv 2505.24641 v1 pith:OCWJZJKF submitted 2025-05-30 cs.CV

Chengzhi Wu , Qianliang Huang , Kun Jin , Julius Pfrommer , Jürgen Beyerer This is my paper · ORCID
classification cs.CV
keywords pointcloudself-supervisedlearningcontrastivecross-branchfeaturefusioncross-attentionlocal-globalrepresentationBYOL-stylepretraining3DShapeNet
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

Existing contrastive learning frameworks compare features only after each branch has processed its input independently; the paper argues this wastes information that could be exchanged earlier. It proposes PoCCA, a self-supervised point-cloud framework that splits the online and target branches into global and local sub-branches, aligns local patch features across the two augmented views, and fuses them into the global features through cross-attention before the contrastive loss is computed. Across linear and fine-tuned classification, few-shot learning, and part segmentation, the pretrained encoder beats all point-cloud-only self-supervised baselines that use no extra training data, with both PointNet and DGCNN backbones. This matters because better single-modal pretraining improves downstream tasks without needing paired images or text.

What carries the argument

The load-bearing device is a four-sub-branch architecture with an aligner and cross-attention fusion. The aligner is a mixture of self-attention and cross-attention over the local patch features of both augmented views, used to put patches from different augmentations into a comparable space; then local-to-global cross-attention treats the global feature as the query and the aligned local patches as key and value, so each branch's global representation absorbs the other branch's local structure before the loss end. Gradient flow follows the online sub-branches, while the target global encoder and target attention module are momentum-updated, and a predictor on online branches prevents collapse. This mechanism is what lets the framework exchange information before the loss, which is the paper's claimed departure from prior contrastive frameworks.

What would settle it

Take the pretrained PoCCA setup with DGCNN and randomly permute the order of the target branch's local patches before they enter the aligner, while keeping the model size and training budget identical; if linear SVM accuracy on ModelNet40 stays essentially at the reported 91.4%, then the aligner's cross-branch correspondence is not doing the work the paper attributes to it.

Watch

Extended reading notes

Core claim

PoCCA establishes that a contrastive learner does not have to wait until the loss to combine information from its branches. By adding sub-branches, the online branch and the target branch each compute a global feature and a set of local patch features; an aligner maps the two sets of patches into a shared space, and local-to-global cross-attention then injects the other branch's local detail into each branch's global representation. The two enriched representations are compared with a symmetrized mean-squared-error loss, with a predictor on the online side and momentum updates on the target side to prevent collapse. After pretraining on ShapeNet, only the encoder is kept, and it outperforms no-extra-data contrastive and pretext baselines on ModelNet40, ScanObjectNN, and ShapeNetPart, and matches or beats several reconstruction-based methods despite using no reconstruction loss.

Load-bearing premise

The load-bearing premise is that the aligner can meaningfully match local patches from two different augmentations of the same point cloud; if it merely mixes unrelated patches, the cross-branch fusion could add noise instead of information.

Editorial extensions

If this is right

  • Pretraining with PoCCA gives linear SVM gains over STRL and other no-extra-data contrastive baselines, with reported 91.4% overall accuracy on ModelNet40 with a DGCNN backbone.
  • Fine-tuning with the pretrained encoder helps most when labels are scarce: with only 1% of ModelNet40 training data, PoCCA plus DGCNN reaches 65.7% accuracy versus 60.5% for STRL.
  • Few-shot classification on ScanObjectNN improves in all reported settings; for example, 5-way 10-shot DGCNN accuracy rises from 74.8% with CrossPoint to 79.9% with PoCCA.
  • Part segmentation on ShapeNetPart with a DGCNN backbone reaches 85.8 instance mIoU, above Jigsaw, OcCo, and the randomly initialized DGCNN baseline.
  • Ablations show that removing local branches, dropping the aligner, or replacing cross-attention with concatenation lowers accuracy, so the specific fusion mechanism drives the reported gains.

Reading between the lines

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

  • Beyond the paper: the same sub-branch exchange idea could be dropped into BYOL-style image or video pipelines, where branches are likewise kept isolated until the loss.
  • Beyond the paper: the aligner's geometric-correspondence assumption could be tested directly by checking whether matched patches from the two augmented views correspond in object coordinates; optimal transport would be a natural alternative to learned attention.
  • Beyond the paper: if alignment is what drives the gain, the framework should transfer to other single-modality 3D inputs such as depth maps or meshes without architectural changes.
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 / 6 minor

Summary. The paper proposes PoCCA, a BYOL-style contrastive self-supervised learning framework for point clouds that does not use extra training data. The key novelty is to allow information exchange between the online and target branches before the contrastive loss is computed. The framework augments an input point cloud into two views, samples global and local patches, uses an aligner module to align local patch features from the two branches, fuses local and global features via cross-attention, and trains the encoder with a symmetric normalized MSE loss. The method is evaluated on ModelNet40 and ScanObjectNN for linear SVM classification, fine-tuned classification, few-shot classification, and on ShapeNetPart for part segmentation, using PointNet, DGCNN, and Transformer backbones. The paper claims state-of-the-art performance among point cloud self-supervised methods that do not use extra training data, and includes ablations of the main architectural components and patch sampling strategies.

Significance. If the empirical results are robust, the cross-branch fusion idea is a useful contribution to point cloud contrastive learning: it goes beyond the standard paradigm of comparing features only at the loss end, and the ablations suggest that both the local sub-branches and the aligner contribute to the gains. The paper also provides a fairly complete set of ablations and compares against a wide range of prior methods. However, the headline claim is not uniformly supported by the paper's own tables: in several settings the proposed method is not strictly better than existing no-extra-data methods, and the main tables lack error bars while the reported margins over strong baselines are often very small. The significance is therefore moderate and contingent on either narrowing the claim or strengthening the evidence. The paper does not mention code release, which limits reproducibility of the empirical claims.

major comments (4)
  1. [Section 1, Section 5, Tables 2 and 4] The unqualified claim that PoCCA achieves state-of-the-art performance among methods that do not use extra training data is not supported by the paper's own comparison tables. In Table 2, fine-tuned classification on ScanObjectNN with the Transformer backbone, Point-MAE (marked ×, no extra training data) reports 85.2 overall accuracy while PoCCA reports 84.8. In Table 4, few-shot classification on ModelNet40 with DGCNN in the 5-way 20-shot setting, OcCo reports 92.5±1.9 while PoCCA reports 92.1±3.6. In Table 6, PoCCA ties Point-MAE on ShapeNetPart instance mIoU (86.1 vs 86.1). Since the abstract and conclusion state unqualified SOTA, please narrow the claim to 'most settings' or to 'contrastive/pretext-based methods without extra data', or provide additional evidence (e.g., multiple seeds) that supports a universal SOTA claim.
  2. [Section 4.1, Section 4.4, Tables 7 and 8] The hyperparameters of the proposed perception-enlarged multi-scale KNN sampling (scale factor α=0,1,2, patch size K=256, and 8 patches per scale) are selected by comparing test accuracy in Table 8, with no validation set used. Because the final reported numbers are also test accuracy, this selection procedure risks overfitting to the test set. Please report the selection protocol explicitly, or use a validation split for hyperparameter choice and then report test accuracy, so that the improvements are not inflated by test-set selection.
  3. [Section 3.3, Table 7] The aligner is a load-bearing component of the proposed cross-branch fusion, but the paper does not validate that the aligned local patch features actually correspond to the same geometric regions across the two augmentations. The only direct evidence is the ablation in Table 7, where removing the aligner drops accuracy from 91.4 to 89.7. This shows the module helps empirically, but not that alignment is the reason. Please add a diagnostic experiment, for example measuring patch-correspondence accuracy between the aligned features, or replacing the aligner's cross-attention with a random permutation of patch indices, to demonstrate that the alignment mechanism itself is responsible for the gain rather than the extra parameters.
  4. [Tables 1, 2, and 6] The main classification and segmentation results are reported as single-run accuracies without error bars or significance tests. The margins over strong no-extra-data baselines are often at or below 0.3 points, e.g., Table 2 DGCNN on ModelNet40 (PoCCA 93.2 vs OcCo 93.0), Table 1 DGCNN on ModelNet40 (PoCCA 91.4 vs STRL 90.9), and Table 6 Transformer instance mIoU (PoCCA 86.1 vs Point-MAE 86.1). For claims of state-of-the-art performance, please report mean and standard deviation over at least three seeds for the main tables, or otherwise demonstrate that the differences are not within run-to-run noise.
minor comments (6)
  1. [References] Reference [49] is cited as the source of BYOL, but the listed entry is Richemond et al., 'BYOL works even without batch statistics'; the original BYOL paper by Grill et al. appears to be missing or mis-cited. Please correct this citation.
  2. [References] References [34] and [35] are the same Stratified Transformer paper, and references [16] and [17] appear to be the same self-contrastive learning paper with different venue strings. Please deduplicate these entries.
  3. [Section 4.3, Table 5] The text for the ScanObjectNN few-shot results says the accuracy 'is also increased significantly in most settings', but the caption of Table 5 says 'in all the reported settings'. Please align the wording between the text and the caption.
  4. [Table 6] For Point-MAE, the category mIoU is shown as '-' while instance mIoU is 86.1. If the category mIoU was not reported in the original paper, please say so in the caption rather than using a dash that could be interpreted as a missing value.
  5. [Section 3.3, Figure 3] Figure 3 is referenced in the description of the aligner, but the figure panels (self-attention, cross-attention, classical cross-attention, offset attention) are not explained in the text. Please add a short description of what each panel shows.
  6. [Sections 3.2 and 3.4] The description of which encoder is used on the target patch branch is inconsistent: Section 3.2 states that the target patch branch shares f_θ1 with the online branch, while Section 3.4 refers to 'the target branch encoder f_θ2' for processing the second augmentation. Please clarify which encoder (and which momentum update rule) applies to the target global branch versus the target patch branch in the loss computation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the results are empirical benchmark comparisons, and the only self-citation is a descriptive related-work mention that is not load-bearing.

full rationale

The paper's central claim is an empirical benchmark comparison, not a derivation from fitted constants. The PoCCA objective (Eq. 3) is a standard BYOL-style normalized mean-square-error loss between online and target branch outputs, and the downstream evaluations either fit a linear SVM on frozen features or fine-tune on held-out labeled splits against externally published baselines. No parameter is fitted to the evaluation target and then reported as a prediction. The multi-scale factor alpha and related hyperparameters are chosen through ablations on the same benchmark, which is a possible test-set selection concern, but it is a mild methodological issue rather than a definitional circularity. The only self-citation, Wu et al. [60] in Section 2.3, is a descriptive related-work mention about attention-based point cloud edge sampling and does not justify any load-bearing architectural choice or result. The abstract's unqualified claim of state-of-the-art is broader than some of the paper's own tables support (e.g., Table 2 shows Point-MAE at 85.2 vs PoCCA at 84.8 on ScanObjectNN, and Table 4 shows OcCo at 92.5 +/- 1.9 vs PoCCA at 92.1 +/- 3.6 in one few-shot setting), but overclaiming relative to one's own comparison tables is a correctness or calibration issue, not circularity. The aligner's geometric-correspondence assumption is also not independently validated, but the aligner is trained end-to-end and not defined in terms of the downstream metric. No load-bearing step reduces by construction to its own input, so the appropriate finding is no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 3 invented entities

PoCCA's central claim is empirical: adding pre-loss cross-branch fusion improves point cloud self-supervised learning. The framework introduces several new components (aligner, cross-attention fusion, multi-scale patch sampling) whose efficacy is demonstrated only through the paper's own experiments. There are no parameter-free derivations; all hyperparameters are chosen by hand or via test-set validation, which slightly increases the circularity burden. The axioms are standard SSL assumptions plus one paper-specific assumption about the aligner's ability to align features across augmentations.

free parameters (6)
  • EMA decay tau = 0.99
    Chosen by hand for the momentum update of the target branch, following BYOL and MoCo.
  • Multi-scale factor alpha = 0, 1, 2
    Selected based on test accuracy in the ablation study (Table 8); this is a mild form of fitting to the benchmark.
  • Number of patches per scale = 8
    Chosen by hand; controls the number of local patches sampled in the local sub-branch.
  • Patch size K = 256
    Chosen by hand; determines the number of points in each local patch.
  • Learning rate and weight decay = 1e-4 each
    Standard Adam optimizer settings chosen by hand.
  • Predictor architecture = MLP with two linear layers and batch norm
    Borrowed from BYOL; the number of hidden units is not specified in the paper.
assumptions (4)
  • domain assumption ShapeNet pretraining transfers to downstream benchmarks (ModelNet40, ScanObjectNN, ShapeNetPart).
    The authors pre-train on ShapeNet and evaluate on other datasets, assuming the learned representations are transferable.
  • domain assumption Two augmentations of the same input point cloud share the same semantic content, making them valid positive pairs.
    Standard assumption in contrastive learning, inherited from SimCLR/BYOL.
  • ad hoc to paper The aligner can align local patch features from different augmentations so that cross-attention fusion is meaningful.
    This is specific to PoCCA and is only validated indirectly through ablations (Table 7).
  • domain assumption Using a predictor and momentum update prevents collapse, as in BYOL.
    Borrowed from BYOL; the authors verify that removing the predictor causes collapse (Table 7, rows with 8.3% and 7.7% accuracy).
invented entities (3)
  • Aligner module
    purpose: Aligns local patch features from two augmentations before cross-attention fusion
    The only evidence for its utility is the internal ablation (Table 7); no independent external verification.
  • Cross-branch local-global cross-attention fusion
    purpose: Fuses global features from one branch with local features from the other branch before the contrastive loss
    Novel component; validation is limited to this paper's experiments.
  • Perception-enlarged multi-scale KNN patch sampling
    purpose: Generates local patches with multi-scale receptive fields by gathering 2^alpha K neighbors and downsampling to K points with FPS
    Introduced as a heuristic; its advantage is shown only in the paper's ablation (Table 8).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Cross Branch Fusion-Based Contrastive Learning Framework for Point Cloud Self-supervised Learning." pith.science (2026). https://pith.science/paper/OCWJZJKF

@misc{pith2026250524641,
  author       = {Pith},
  title        = {Pith review of: A Cross Branch Fusion-Based Contrastive Learning Framework for Point Cloud Self-supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OCWJZJKF}},
  note         = {Machine review of arXiv:2505.24641}
}
read the original abstract

Contrastive learning is an essential method in self-supervised learning. It primarily employs a multi-branch strategy to compare latent representations obtained from different branches and train the encoder. In the case of multi-modal input, diverse modalities of the same object are fed into distinct branches. When using single-modal data, the same input undergoes various augmentations before being fed into different branches. However, all existing contrastive learning frameworks have so far only performed contrastive operations on the learned features at the final loss end, with no information exchange between different branches prior to this stage. In this paper, for point cloud unsupervised learning without the use of extra training data, we propose a Contrastive Cross-branch Attention-based framework for Point cloud data (termed PoCCA), to learn rich 3D point cloud representations. By introducing sub-branches, PoCCA allows information exchange between different branches before the loss end. Experimental results demonstrate that in the case of using no extra training data, the representations learned with our self-supervised model achieve state-of-the-art performances when used for downstream tasks on point clouds.

Figures

Figures reproduced from arXiv: 2505.24641 by the authors.

Figure 1
Figure 1. The framework of our proposed PoCCA. Given an input point cloud, it is first augmented with two different augmentation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Perception enlarged KNN-based patch sampling. Af [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Detailed architecture of the aligner module and two [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of features on the test split of Mod [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 58 canonical work pages

  1. [1]

    Learning representations and generative models for 3D point clouds

    Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3D point clouds. InProceedings of the 35th In- ternational Conference on Machine Learning, pages 40–49,

    work page

  2. [2]

    Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Ro- drigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9892–9902, 2022. 1, 2, 6, 7, 8

    work page 2022

  3. [3]

    Beit: Bert pre-training of image transformers.ArXiv, abs/2106.08254, 2022

    Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers.ArXiv, abs/2106.08254, 2022. 1

    arXiv 2022

  4. [4]

    Czar- necki

    Prarthana Bhattacharyya, Chengjie Huang, and K. Czar- necki. Sa-det3d: Self-attention based context-aware 3d ob- ject detection.2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 3022–3031,

    work page 2021

  5. [5]

    Large scale gan training for high fidelity natural image synthesis

    Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. ArXiv, abs/1809.11096, 2019. 1

    arXiv 2019

  6. [6]

    Unsupervised learning of visual features by contrasting cluster assignments.ArXiv, abs/2006.09882, 2020

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Pi- otr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments.ArXiv, abs/2006.09882, 2020. 2

    arXiv 2006

  7. [7]

    Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu

    Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Mano- lis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repos- itory.arXiv:1512.03012, 2015. 5, 7

    arXiv 2015

  8. [8]

    Crossvit: Cross-attention multi-scale vision transformer for image classification.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 347–356, 2021

    Chun-Fu Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 347–356, 2021. 4

    work page 2021

Show all 75 references
  1. [9]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey E. Hinton. A simple framework for contrastive learn- ing of visual representations.ArXiv, abs/2002.05709, 2020. 1, 2

    2002 arXiv

  2. [10]

    Exploring simple siamese rep- resentation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese rep- resentation learning. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 15750–15758, 2021. 1, 2, 4, 5

    2021

  3. [11]

    Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and P. Abbeel. Infogan: Interpretable represen- tation learning by information maximizing generative adver- sarial nets. InNIPS, 2016. 1

    2016

  4. [12]

    Shape self- correction for unsupervised point cloud understanding.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8362–8371, 2021

    Ye Chen, Jinxian Liu, Bingbing Ni, Hang Wang, Jiancheng Yang, Ning Liu, Teng Li, and Qi Tian. Shape self- correction for unsupervised point cloud understanding.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8362–8371, 2021. 6

    2021

  5. [13]

    Patchformer: A versatile 3d transformer based on patch at- tention.ArXiv, abs/2111.00207, 2021

    Zhang Cheng, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: A versatile 3d transformer based on patch at- tention.ArXiv, abs/2111.00207, 2021. 3

    2021 arXiv

  6. [14]

    Cy- clegan, a master of steganography.ArXiv, abs/1712.02950,

    Casey Chu, Andrey Zhmoginov, and Mark Sandler. Cy- clegan, a master of steganography.ArXiv, abs/1712.02950,

    arXiv

  7. [15]

    An image is worth 16x16 words: Transformers for image recognition at scale.ArXiv, abs/2010.11929, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2010 arXiv

  8. [16]

    Self-contrastive learning with hard negative sampling for self-supervised point cloud learning

    Bi’an Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning. InACM Multimedia Conference, pages 3133–3142, 2021. 6

    2021

  9. [17]

    Self-contrastive learning with hard negative sampling for self-supervised point cloud learning.Proceedings of the 29th ACM Inter- national Conference on Multimedia, 2021

    Bianli Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning.Proceedings of the 29th ACM Inter- national Conference on Multimedia, 2021. 1, 6, 8

    2021

  10. [18]

    Benjamin Eckart, Wentao Yuan, Chao Liu, and Jan Kautz. Self-supervised learning on 3d point clouds by learning dis- crete generative models.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8244–8253, 2021. 6

    2021

  11. [19]

    Yuval Eldar, Michael Lindenbaum, Moshe Porat, and Yehoshua Y . Zeevi. The farthest point strategy for pro- gressive image sampling.Proceedings of the 12th IAPR International Conference on Pattern Recognition, Vol. 2 - Conference B: Computer Vision & Image Processing. (Cat. No....

    1994

  12. [20]

    Nico Engel, Vasileios Belagiannis, and Klaus C. J. Diet- mayer. Point transformer.IEEE Access, 9:134826–134840,

  13. [21]

    Support vec- tor machines: Theory and applications

    Theodoros Evgeniou and Massimiliano Pontil. Support vec- tor machines: Theory and applications. InMachine Learning and Its Applications, 2001. 6

    2001

  14. [22]

    Object classification from a single example utilizing class relevance metrics

    Michael Fink. Object classification from a single example utilizing class relevance metrics. InNIPS, 2004. 7

    2004

  15. [23]

    Mul- tiresolution tree networks for 3D point cloud processing

    Matheus Gadelha, Rui Wang, and Subhransu Maji. Mul- tiresolution tree networks for 3D point cloud processing. In ECCV, 2018. 6

    2018

  16. [24]

    Pct: Point cloud transformer.Comput

    Meng-Hao Guo, Junxiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph Robert Martin, and Shimin Hu. Pct: Point cloud transformer.Comput. Vis. Media, 7:187–199, 2021. 3, 6, 8

    2021

  17. [25]

    Zhizhong Han, Mingyang Shang, Yu-Shen Liu, and Matthias Zwicker. View inter-prediction gan: Unsupervised represen- tation learning for 3D shapes by learning global shape mem- ories to support local view predictions.Proceedings of the AAAI Conference on Artificial Intelligence,...

  18. [26]

    Unsupervised multi-task feature learning on point clouds

    Kaveh Hassani and Mike Haley. Unsupervised multi-task feature learning on point clouds. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6

    2019

  19. [27]

    Girshick

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross B. Girshick. Momentum contrast for unsupervised vi- sual representation learning.2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9726–9735, 2020. 1, 2

    2020

  20. [28]

    Girshick

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Pi- otr Doll’ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988, 2022. 1

    2022

  21. [29]

    Spatio-temporal self-supervised representation learning for 3d point clouds.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6515–6525, 2021

    Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6515–6525, 2021. 1, 2, 6, 7

    2021

  22. [30]

    Pyramid point cloud transformer for large-scale place recog- nition.2021 IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 6078–6087, 2021

    Le Hui, Hang Yang, Mingmei Cheng, Jin Xie, and Jian Yang. Pyramid point cloud transformer for large-scale place recog- nition.2021 IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 6078–6087, 2021. 3

    2021

  23. [31]

    Longlong Jing, Yucheng Chen, Ling Zhang, Mingyi He, and Yingli Tian. Self-supervised feature learning by cross- modality and cross-view correspondences.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1581–1891, 2020. 8

    2021

  24. [32]

    Progressive growing of gans for improved quality, stability, and variation.ArXiv, abs/1710.10196, 2018

    Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation.ArXiv, abs/1710.10196, 2018. 1

    2018 arXiv

  25. [33]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. 2019 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 4396–4405, 2019. 2

    2019

  26. [34]

    Stratified trans- former for 3d point cloud segmentation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8490–8499, 2022

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified trans- former for 3d point cloud segmentation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8490–8499, 2022. 4

    2022

  27. [35]

    Strati- fied transformer for 3d point cloud segmentation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 8490–8499, 2022

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Strati- fied transformer for 3d point cloud segmentation.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR), pages 8490–8499, 2022. 3

    2022

  28. [36]

    Li, Matthias M ¨uller, Ali K

    G. Li, Matthias M ¨uller, Ali K. Thabet, and Bernard Ghanem. Deepgcns: Can gcns go as deep as cnns?2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9266–9275, 2019. 4

    2019

  29. [37]

    Chen, and Gim Hee Lee

    Jiaxin Li, Ben M. Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9397–9406, 2018. 6

    2018

  30. [38]

    Pointcnn: Convolution on x-transformed points

    Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. InAdvances in Neural Information Processing Sys- tems, 2018. 6, 7

    2018

  31. [39]

    Masked discrimi- nation for self-supervised learning on point clouds

    Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimi- nation for self-supervised learning on point clouds. InEuro- pean Conference on Computer Vision, 2022. 2

    2022

  32. [40]

    Relation-shape convolutional neural network for point cloud analysis

    Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895–8904, 2019. 7, 8

    2019

  33. [41]

    3dpct: 3d point cloud transformer with dual self-attention

    Dening Lu, Kyle Gao, Qian Xie, Linlin Xu, and Jonathan Li. 3dpct: 3d point cloud transformer with dual self-attention. ArXiv, abs/2209.11255, 2022. 3

    2022 arXiv

  34. [42]

    3d object detection with pointformer.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7459–7468, 2021

    Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer.2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7459–7468, 2021. 3

    2021

  35. [43]

    Yatian Pang, Wenxiao Wang, Francis E. H. Tay, W. Liu, Yonghong Tian, and Liuliang Yuan. Masked autoencoders for point cloud self-supervised learning. InEuropean Con- ference on Computer Vision, 2022. 2, 6, 8

    2022

  36. [44]

    Omid Poursaeed, Tianxing Jiang, Quintessa Qiao, Nayun Xu, and Vladimir G. Kim. Self-supervised learning of point clouds via orientation estimation.2020 International Con- ference on 3D Vision (3DV), pages 1018–1028, 2020. 2, 6

    2020

  37. [45]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    C. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Point- net: Deep learning on point sets for 3d classification and seg- mentation.2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77–85, 2017. 4, 6, 8

    2017

  38. [46]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. InAdvances in Neural Informa- tion Processing Systems, 2017. 4, 6, 7, 8

    2017

  39. [47]

    Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining.ArXiv, abs/2302.02318, 2023

    Zekun Qi, Runpei Dong, Guo Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining.ArXiv, abs/2302.02318, 2023. 1, 3, 6, 8

    2023 arXiv

  40. [48]

    Unsuper- vised representation learning with deep convolutional gen- erative adversarial networks.CoRR, abs/1511.06434, 2016

    Alec Radford, Luke Metz, and Soumith Chintala. Unsuper- vised representation learning with deep convolutional gen- erative adversarial networks.CoRR, abs/1511.06434, 2016. 1

    2016 arXiv

  41. [49]

    Byol works even without batch statistics.arXiv preprint arXiv:2010.10241, 2020

    Pierre H Richemond, Jean-Bastien Grill, Florent Altch ´e, Corentin Tallec, Florian Strub, Andrew Brock, Samuel Smith, Soham De, Razvan Pascanu, Bilal Piot, et al. Byol works even without batch statistics.arXiv preprint arXiv:2010.10241, 2020. 1, 2, 4, 5

    2010 arXiv

  42. [50]

    Info3d: Representation learning on 3d ob- jects using mutual information maximization and contrastive learning.ArXiv, abs/2006.02598, 2020

    Aditya Sanghi. Info3d: Representation learning on 3d ob- jects using mutual information maximization and contrastive learning.ArXiv, abs/2006.02598, 2020. 1, 6

    2006 arXiv

  43. [51]

    Context prediction for unsupervised deep learning on point clouds.ArXiv, abs/1901.08396, 2019

    Jonathan Sauder and Bjarne Sievers. Context prediction for unsupervised deep learning on point clouds.ArXiv, abs/1901.08396, 2019. 8

    1901 arXiv

  44. [52]

    Self-supervised deep learning on point clouds by reconstructing space.Advances in Neural Information Processing Systems, 32:12962–12972,

    Jonathan Sauder and Bjarne Sievers. Self-supervised deep learning on point clouds by reconstructing space.Advances in Neural Information Processing Systems, 32:12962–12972,

  45. [53]

    Self-supervised few-shot learning on point clouds

    Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. InAdvances in Neural Information Processing Systems, pages 7212–7221, 2020. 7

    2020

  46. [54]

    Con- trastive multiview coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Con- trastive multiview coding. InECCV, 2020. 2

    2020

  47. [55]

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

    Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1588–1597,

  48. [56]

    Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N

    Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need.ArXiv, abs/1706.03762, 2017. 3

    2017 arXiv

  49. [57]

    Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matthew J. Kusner. Unsupervised point cloud pre-training via occlusion completion. InInternational Conference on Computer Vision, ICCV, 2021. 2, 6, 7, 8

    2021

  50. [58]

    Sarma, Michael M

    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), 38:1 – 12, 2019. 4, 6, 8

    2019

  51. [59]

    Learning generalizable part-based feature representation for 3d point clouds

    Xin Wei, Xiang Gu, and Jian Sun. Learning generalizable part-based feature representation for 3d point clouds. InNeu- ral Information Processing Systems, 2022. 2

    2022

  52. [60]

    Attention-based point cloud edge sampling

    Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and J ¨urgen Beyerer. Attention-based point cloud edge sampling. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5333–5343, 2023. 3

    2023

  53. [61]

    Learning a probabilistic latent space of ob- ject shapes via 3D generative-adversarial modeling

    Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of ob- ject shapes via 3D generative-adversarial modeling. InAd- vances in Neural Information Processing Systems, 2016. 7

    2016

  54. [62]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin- guang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912–1920, 2015. 6, 7

    2015

  55. [63]

    Yu, and Dahua Lin

    Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination.2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3733–3742, 2018. 2

    2018

  56. [64]

    Qi, Leonidas J

    Saining Xie, Jiatao Gu, Demi Guo, C. Qi, Leonidas J. Guibas, and Or Litany. Pointcontrast: Unsupervised pre-training for 3d point cloud understanding.ArXiv, abs/2007.10985, 2020. 1, 2

    2007 arXiv

  57. [65]

    Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9643–9653, 2022

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: a simple framework for masked image modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9643–9653, 2022. 1

    2022

  58. [66]

    Fold- ingnet: Point cloud auto-encoder via deep grid deformation

    Yaoqing Yang, Chen Feng, Yiru Shen, and Dong Tian. Fold- ingnet: Point cloud auto-encoder via deep grid deformation. 2018 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 206–215, 2018. 5, 6, 7

    2018

  59. [67]

    Yi, Vladimir G

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

    2016

  60. [68]

    Pointr: Diverse point cloud comple- tion with geometry-aware transformers.2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 12478–12487, 2021

    Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud comple- tion with geometry-aware transformers.2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 12478–12487, 2021. 3

    2021

  61. [69]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291–19300, 2021

    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.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19291–19300, 2021. 2, 6, 8

    2022

  62. [70]

    Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks

    Ling Zhang and Zhigang Zhu. Unsupervised feature learning for point cloud understanding by contrasting and clustering using graph convolutional neural networks. InInternational Conference on 3D Vision (3DV), pages 395–404, 2019. 6

    2019

  63. [71]

    Point-m2ae: Multi-scale masked autoencoders for hierarchi- cal point cloud pre-training.ArXiv, abs/2205.14401, 2022

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bingyan Zhao, Dong Lei Wang, Yu Jiao Qiao, and Hongsheng Li. Point-m2ae: Multi-scale masked autoencoders for hierarchi- cal point cloud pre-training.ArXiv, abs/2205.14401, 2022. 3, 6, 8

    2022 arXiv

  64. [72]

    Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders

    Renrui Zhang, Liuhui Wang, Yu Jiao Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. ArXiv, abs/2212.06785, 2022. 1, 3, 6, 8

    2022 arXiv

  65. [73]

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip H. S. Torr, and Vladlen Koltun. Point transformer.2021 IEEE/CVF In- ternational Conference on Computer Vision (ICCV), pages 16239–16248, 2021. 3

    2021

  66. [74]

    3D point capsule networks

    Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3D point capsule networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6, 7

    2019

  67. [75]

    Seedformer: Patch seeds based point cloud completion with upsample transformer

    Hao Zhou, Yun Cao, Wenqing Chu, Junwei Zhu, Tong Lu, Ying Tai, and Chengjie Wang. Seedformer: Patch seeds based point cloud completion with upsample transformer. ArXiv, abs/2207.10315, 2022. 3

    2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.