Pith. sign in

REVIEW 4 major objections 6 minor 89 references

The paper claims that cross-reconstruction between two decoupled views—randomly cropped, normalized, and rotated—is a more informative generative pre-training task for point clouds than masked self-reconstruction, and that the resulting fro

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 12:42 UTC pith:2MKKAM5D

load-bearing objection A genuinely new two-view cross-reconstruction recipe for point-cloud SSL with big empirical gains, but the paper's own equations undermine the claimed geometric meaning of its core position encoding; worth refereeing, needs a major revision. the 4 major comments →

arxiv 2509.01250 v1 pith:2MKKAM5D submitted 2025-09-01 cs.CV

Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views

Xiangdong Zhang , Shaofeng Zhang , Junchi Yan This is my paper
classification cs.CV
keywords self-supervised learningpoint cloudcross-reconstructionmasked autoencoderdecoupled viewsview-relative positional embeddingpositional query3D representation learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Point-PQAE is built on a single claim: for point-cloud self-supervised learning, the right generative task is not 'recover what was masked' but 'reconstruct one view from another view that has been made genuinely different.' The paper argues that masked self-reconstruction within one view is too easy, because the visible and missing parts share one coordinate frame and one normalization; cross-reconstruction between two cropped, independently normalized and rotated views forces the model to learn both the shape of a view and where one view sits relative to the other. On the paper's evidence, this harder task pays off: frozen encoder features from Point-PQAE beat the Point-MAE baseline by 6.5%, 7.0%, and 6.7% on the three ScanObjectNN variants under a linear-probe protocol (MLP-LINEAR), with similar gains in few-shot learning. If the claim holds, it gives 3D practitioners a new label-free training signal that needs no 2D teachers or multi-modal data.

Core claim

The central discovery is that the difficulty of the pre-training task, not the architecture, is what drives representation quality. Point-PQAE generates two decoupled views of the same point cloud by randomly selecting two center points, cropping the r1·p and r2·p nearest points around each, then min-max normalizing each crop around its own geometric center and applying an independent random rotation. It encodes both views with a shared Transformer and, for each direction, uses the view-relative positional embedding (VRPE)—a fixed sinusoid over the target view's patch centers concatenated with the difference of the two geometric centers—as the query in a cross-attention layer over the source

What carries the argument

The load-bearing mechanism is a positional query block between encoder and decoder. VRPE is a fixed sinusoidal positional encoding of the 6-dimensional vector formed by concatenating the target view's patch-group centers with the difference of the two views' geometric centers (L1−L2); it is used as the query Q in cross-attention whose keys and values come from the encoded source view, so the decoder gets an explicit hint about where the target view lies relative to the source. Two supporting mechanisms carry the rest: a random-crop view generator that selects two overlapping neighborhoods around random center points and then decouples them via min-max normalization and independent random rot

Load-bearing premise

The method assumes that the position signal derived from the two views' original centers still guides reconstruction correctly, even after each view has been independently rotated and normalized; the paper does not test this mismatch.

What would settle it

Rotate the relative-position vector L1−L2 by the same rotation matrices applied to the two cropped views before constructing the positional query, and rerun the ScanObjectNN MLP-LINEAR evaluation. If accuracy does not drop, the current VRPE is not encoding true inter-view geometry and the claimed mechanism is mis-specified; if accuracy drops, the unrotated signal is load-bearing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Frozen features transfer better: on the three ScanObjectNN variants, MLP-LINEAR accuracy rises by 6.5, 7.0, and 6.7 points over the self-reconstruction baseline with the same 22.1M-parameter backbone.
  • Few-shot classification improves, especially under the linear-probe protocol: e.g., 10-way 20-shot goes from 89.7 on the Point-MAE baseline to 93.5.
  • Because the positional query block is a separable module, the paper shows it can be grafted onto knowledge distillation, letting a student recover a teacher's representation of a second view from a first view.
  • The symmetric siamese loss contributes beyond one-directional reconstruction; ablations show it adds accuracy on all three ScanObjectNN variants.
  • Qualitative results show cross-reconstruction generalizes to crop ratios other than the 0.6 minimum used in pre-training, suggesting the learned inter-view relations are not tied to a single crop scale.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: the VRPE is built from view centers recorded before independent rotations and normalization, so after those augmentations the true relative offset is a rotated, rescaled version of L1−L2. A natural next test is to rotate the relative vector by the same rotations applied to the two views before encoding; if accuracy holds or improves, the paper's mechanism is mis-specified, and if it dro
  • Inference: the view-generation scheme is augmentation-based and the query module is detached from any specific downstream head, so the same decoupled-view recipe could be dropped into contrastive or distillation objectives, where two views already exist, and into dense prediction tasks such as detection that require partial-view reasoning.
  • Inference: the minimum crop ratio rm acts as a difficulty dial. The paper finds 0.6 optimal for ScanObjectNN classification; lower ratios should make reconstruction harder and may expose finer local geometry that helps dense downstream tasks like part or scene segmentation, at some cost in convergence.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces Point-PQAE, a self-supervised pre-training method for point clouds based on cross-reconstruction between two decoupled views. Two views are generated by random cropping, min-max normalization around each crop's geometric center, and per-view random rotation. A shared encoder processes both views; a proposed View-Relative Positional Embedding (VRPE) is concatenated with patch centers and used as the query in a cross-attention block to reconstruct one view from the other. Pre-training on ShapeNet is transferred to ScanObjectNN, ModelNet40, few-shot classification, ShapeNetPart, and S3DIS. The paper reports large gains over Point-MAE, especially under the MLP-LINEAR protocol on ScanObjectNN.

Significance. If the mechanism worked as described, this would be a valuable demonstration that cross-view reconstruction is more informative than masked self-reconstruction for 3D representation learning. The paper ships code, evaluates on standard external benchmarks, and includes a fairly complete ablation set covering crop ratio, augmentation, loss function, siamese loss, and positional-embedding variants. The strengths are the empirical breadth and the clear separation of cross-reconstruction from self-reconstruction. However, the central geometric claim about VRPE is questionable, and the empirical claims are partly overstated relative to the paper's own appendix results.

major comments (4)
  1. [Sec. 3.1, Eqs. (8)-(10)] VRPE mixes coordinate frames. L1 and L2 are recorded before min-max normalization and per-view random rotation, whereas G1 and G2 are computed after those transformations. For a corresponding point, the actual offset between the transformed views is an affine function of L1-L2, the per-view scales, and the rotations, not simply L1-L2. Therefore RP_{1->2} = ConCat(G2, L1-L2) is not a view-relative position embedding in the geometric sense claimed in the abstract and Sec. 3.2. This undermines the interpretation of the ablation in Table 4a and the conceptual novelty. The authors should either derive and use the correct transformed relative offset, or explicitly test whether the mismatched query is harmless, or revise the claimed geometric meaning.
  2. [Sec. 4.3, Table 4a] The VRPE design confounds 'relative position between views' with 'target patch centers G2'. Since RP_{1->2} includes G2, the improvement of VRPE over the None condition could come from providing the target patch locations rather than from any relative geometric relation between the two views. Add an ablation with G2 only (e.g., ConCat(G2, 0)) and one with the correctly transformed relative offset, to isolate the contribution actually claimed. Without this, the paper's interpretation of the 95.0/93.6/89.6 result is not supported.
  3. [Sec. 4.2 vs Appendix Table 6] The main text claims Point-PQAE 'surpasses previous single-modal self-supervised methods' on ScanObjectNN, but Appendix Table 6 shows PCP-MAE, a self-reconstruction method, achieving 95.5/94.3/90.4 on OBJ-BG/OBJ-ONLY/PB-T50-RS under FULL, versus 95.0/93.6/89.6 for Point-PQAE. The claim is also contradicted on ModelNet40 1K (94.2 vs 93.9). Please revise the positioning statements and include recent self-reconstruction baselines in the main comparison or explicitly explain the discrepancy.
  4. [Sec. 4.3, 'Analysis on decoupled views generation'] The text acknowledges that the cropped views can overlap, and crop ratios are sampled from [0.6,1]. This means many pairs have large overlap, so after normalization and rotation the task may reduce partly to aligning augmented versions of nearly identical subsets, rather than requiring true cross-view inference. The claim that cross-reconstruction is 'much more difficult' than masked self-reconstruction would be stronger if supported by a measurable difficulty signal: training loss curves, overlap statistics, or a controlled experiment bounding the overlap ratio.
minor comments (6)
  1. [Introduction] Typo: 'pertaining' should be 'pre-training' in the first paragraph.
  2. [Sec. 3.2, Eq. (10)] The indexing in Eq. (10) is unclear: RPi_{1->2} is used both as an n-row matrix and as a scalar sinusoid argument. Please specify that the sinusoid is applied per dimension.
  3. [Sec. 4.2, Part segmentation] Duplicate sentence: 'The results are reported in Tab. 3a, The results are reported in Tab. 3a'.
  4. [Table 4 caption] The caption says default settings are marked in gray, but the table as rendered does not show gray highlighting.
  5. [Appendix Fig. 4] The x-axis label 'Query Crop ratio' should be 'Minimum crop ratio rm'.
  6. [Table 1 vs Appendix Table 6] Point-PQAE's ModelNet40 accuracy is reported as 94.0/94.3 in Table 1 but 93.9/94.3 in Appendix Table 6. Please reconcile.

Circularity Check

0 steps flagged

No significant circularity: the central claims are external empirical comparisons; the VRPE frame-mismatch and ablation-on-benchmark are correctness/evaluation risks, not circular derivations.

full rationale

This is an empirical method paper, not a derivation chain, so the classic circularity failure modes do not apply. The headline claim—that Point-PQAE's cross-reconstruction improves SSL for point clouds—is tested by pre-training on ShapeNet and transferring to external benchmarks (ScanObjectNN, ModelNet40, ShapeNetPart, S3DIS) whose labels are not used to fit the pre-training objective. No parameter is fitted to a subset of these benchmarks and then reported as a prediction of a closely related quantity; the MLP-LINEAR and MLP-3 evaluations are standard linear-probe protocols with frozen features. The VRPE construction in Sec. 3.1, Eq. 8-10 does have a real internal-consistency problem: L1 and L2 are recorded before per-view min-max normalization and random rotation, while G1 and G2 live in the transformed local frames, so RL1->2 = L1 - L2 is not the true relative offset in the normalized/rotated coordinate systems. However, this is a correctness/geometry issue, not circularity—the paper does not define VRPE in terms of the target reconstruction and then claim the reconstruction as evidence; the claimed geometric meaning is unsupported, but not equivalent to its inputs by construction. Likewise, the many self-citations (e.g., PCP-MAE [87], PQCL/PQDiff [84,86]) are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from same-author prior work to force the central design. Finally, the ablation tables (Tab. 4, Fig. 4) select default components on ScanObjectNN, the same benchmark used in the headline comparison; this is a benchmark-selection risk and should be noted, but the reported improvements over Point-MAE are still independent external comparisons and not logically forced by a fitted equation. Overall, no circular step is exhibited in the paper's own derivation or evaluation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 3 invented entities

All central components are new design choices whose evidence is internal (ablations on ScanObjectNN). The paper has no free parameters in the analytical sense, but several hyperparameters were tuned on the benchmark used for the headline claims, and the rotation/VRPE mismatch is an unstated premise. The invented entities are method components, not physical postulates, and their only evidence is the paper's own reported results.

free parameters (5)
  • minimum crop ratio rm = 0.6
    Selected by ablation on ScanObjectNN (Fig. 4). Controls view overlap and task difficulty.
  • view augmentation type = rotation
    Chosen among jitter, scale, rotation, scale&translate, rotation+scale&translate via Tab. 4b on ScanObjectNN.
  • reconstruction loss = CD-l2
    Chosen among cosine, CD-l1, CD-l2 via Tab. 9 on ScanObjectNN.
  • siamese loss structure = L2->1 + L1->2
    Ablated against the one-directional loss in Tab. 10 on ScanObjectNN.
  • VRPE encoding type = sinusoid (Eq. 10)
    Chosen among None, APE, learnable, sinusoid via Tab. 4a on ScanObjectNN.
axioms (4)
  • domain assumption Cropping by nearest points around a random center yields meaningful object views for SSL
    Core premise of the view generation; introduced in Sec. 3.1.
  • domain assumption Pre-training on ShapeNet transfers to ScanObjectNN, ModelNet40, ShapeNetPart, S3DIS
    Standard SSL transfer assumption used for all downstream claims.
  • ad hoc to paper The relative offset L1-L2 recorded before rotation remains a valid positional query after independent random rotations and normalizations
    Unstated in Sec. 3.2; the actual relative geometry after augmentation differs from the encoded one.
  • domain assumption Benchmark accuracy on ScanObjectNN is a reliable proxy for representation quality
    Used for both ablations and headline claims.
invented entities (3)
  • View-Relative Positional Embedding (VRPE) no independent evidence
    purpose: Encodes relative position between two decoupled views to query the target view from the source view.
    Supported only by in-paper ablations (Tab 4a); no external validation outside this paper.
  • Positional Query block (PQ) no independent evidence
    purpose: Cross-attention using VRPE as query and source-view features as K/V to reconstruct the other view.
    Same, internal ablation evidence only.
  • 3D random crop mechanism no independent evidence
    purpose: Generates two decoupled views by nearest-point cropping around random centers.
    Internal ablations only; the novelty claim relative to 2D crop is not independently verified.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 23502 in / 11685 out tokens · 133524 ms · 2026-08-05T12:42:26.521876+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views." pith.science (2026). https://pith.science/paper/2MKKAM5D

@misc{pith2026250901250,
  author       = {Pith},
  title        = {Pith review of: Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MKKAM5D}},
  note         = {Machine review of arXiv:2509.01250}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by 6.5%, 7.0%, and 6.7% in three variants of ScanObjectNN with the Mlp-Linear evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE.

Figures

Figures reproduced from arXiv: 2509.01250 by Junchi Yan, Shaofeng Zhang, Xiangdong Zhang.

Figure 1
Figure 1. Figure 1: Comparison between self-reconstruction (Point [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Pipeline of Point-PQAE. The input point cloud is randomly cropped followed by the rotation to generate views. Then, we feed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Cross-reconstruction results on ShapeNet. The arrow points from source point clouds to cross-reconstruction results. Point [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study on different minimum crop ratios [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 64 canonical work pages · 3 internal anchors

  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. In International conference on machine learning, pages 40–49. PMLR, 2018. 1

    2018

  2. [2]

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

    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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902–9912, 2022. 1, 2, 7, 8

    2022

  3. [3]

    3d seman- tic parsing of large-scale indoor spaces

    Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese. 3d seman- tic parsing of large-scale indoor spaces. In Proceedings of 13 the IEEE conference on computer vision and pattern recog- nition, pages 1534–1543, 2016. 6

    2016

  4. [4]

    Beit: Bert pre-training of image transformers

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 3, 7

    Pith/arXiv arXiv 2021

  5. [5]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. In Pro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 1

    2021

  6. [6]

    Shapenet: An information-rich 3d model repository

    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. 5, 11

    Pith/arXiv arXiv 2015

  7. [7]

    Pimae: Point cloud and image interactive masked autoencoders for 3d object detection

    Anthony Chen, Kevin Zhang, Renrui Zhang, Zihan Wang, Yuheng Lu, Yandong Guo, and Shanghang Zhang. Pimae: Point cloud and image interactive masked autoencoders for 3d object detection. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5291–5301, 2023. 10

    2023

  8. [8]

    Pointgpt: Auto-regressively generative pre- training from point clouds

    Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre- training from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 5

    2024

  9. [9]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on ma- chine learning, pages 1597–1607. PMLR, 2020. 1, 2, 3, 7

    2020

  10. [10]

    Autoen- coders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022

    Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jian- jian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoen- coders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2, 5, 7, 8, 12

    Pith/arXiv arXiv 2022

  11. [11]

    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. In Proceedings of the 29th ACM In- ternational Conference on Multimedia , pages 3133–3142,

  12. [12]

    A comparative review of hand-eye calibration tech- niques for vision guided robots

    Ikenna Enebuse, Mathias Foo, Babul Salam Ksm Kader Ibrahim, Hafiz Ahmed, Fhon Supmak, and Odongo Steven Eyobu. A comparative review of hand-eye calibration tech- niques for vision guided robots. IEEE Access, 9:113143– 113155, 2021. 1

    2021

  13. [13]

    Efficient image pre-training with siamese cropped masked autoencoders

    Alexandre Eyma ¨el, Renaud Vandeghen, Anthony Cioppa, Silvio Giancola, Bernard Ghanem, and Marc Van Droogen- broeck. Efficient image pre-training with siamese cropped masked autoencoders. In European Conference on Computer Vision, pages 348–366. Springer, 2025. 9, 11

    work page 2025

  14. [14]

    A point set generation network for 3d object reconstruction from a single image

    Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605–613, 2017. 5, 12

    work page 2017

  15. [15]

    Revisiting point cloud shape classification with a simple and effective baseline

    Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning, pages 3809–3820. PMLR, 2021. 7

    work page 2021

  16. [16]

    Bootstrap your own latent-a new approach to self-supervised learning

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Ghesh- laghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271–21284, 2020. 1, 2, 7

    work page 2020

  17. [17]

    Pct: Point cloud transformer

    Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187–199, 2021. 1, 7

    work page 2021

  18. [18]

    Joint-mae: 2d-3d joint masked au- toencoders for 3d point cloud pre-training

    Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked au- toencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023. 7, 8, 9, 10, 12

    Pith/arXiv arXiv 2023

  19. [19]

    Siamese masked autoencoders

    Agrim Gupta, Jiajun Wu, Jia Deng, and Fei-Fei Li. Siamese masked autoencoders. Advances in Neural Information Pro- cessing Systems, 36:40676–40693, 2023. 9, 11

    work page 2023

  20. [20]

    Dynamic focus-aware po- sitional queries for semantic segmentation

    Haoyu He, Jianfei Cai, Zizheng Pan, Jing Liu, Jing Zhang, Dacheng Tao, and Bohan Zhuang. Dynamic focus-aware po- sitional queries for semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11299–11308, 2023. 11

    work page 2023

  21. [21]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 3

    2016

  22. [22]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 9729–9738, 2020. 1, 2, 3

    work page 2020

  23. [23]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 16000– 16009, 2022. 1, 2, 4, 7, 11

    work page 2022

  24. [24]

    Distill- ing the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distill- ing the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 12

    Pith/arXiv arXiv 2015

  25. [25]

    Spatio-temporal self-supervised representation learning for 3d point clouds

    Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 6535–6545,

    work page

  26. [26]

    Imagenet classification with deep convolutional neural net- works

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural net- works. Advances in neural information processing systems , 25, 2012. 3

    work page 2012

  27. [27]

    Semmae: Semantic-guided mask- ing for learning masked autoencoders

    Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided mask- ing for learning masked autoencoders. NeurIPS, 2022. 2

    work page 2022

  28. [28]

    So-net: Self- organizing network for point cloud analysis

    Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self- organizing network for point cloud analysis. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 1

    work page 2018

  29. [29]

    Cross-BERT for Point Cloud Pretraining

    Xin Li, Peng Li, Zeyong Wei, Zhe Zhu, Mingqiang Wei, Jun- hui Hou, Liangliang Nan, Jing Qin, Haoran Xie, and Fu Lee 14 Wang. Cross-bert for point cloud pretraining. arXiv preprint arXiv:2312.04891, 2023. 9, 10

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    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. Advances in neural information processing systems, 31, 2018. 1, 7

    work page 2018

  31. [31]

    Masked dis- crimination for self-supervised learning on point clouds

    Haotian Liu, Mu Cai, and Yong Jae Lee. Masked dis- crimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657–675. Springer, 2022. 1, 2, 7, 8

    work page 2022

  32. [32]

    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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8895– 8904, 2019. 5

    work page 2019

  33. [33]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2

    2021

  34. [34]

    Sgdr: Stochas- tic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochas- tic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5, 12

    Pith/arXiv arXiv 2016

  35. [35]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 , 2017. 5, 12

    Pith/arXiv arXiv 2017

  36. [36]

    An end-to- end transformer model for 3d object detection

    Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to- end transformer model for 3d object detection. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 2906–2917, 2021. 1

    work page 2021

  37. [37]

    Repre- sentation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2

    Pith/arXiv arXiv 2018

  38. [38]

    Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving

    Bo Pang, Hongchi Xia, and Cewu Lu. Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5229–5239, 2023. 7

    work page 2023

  39. [39]

    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 European conference on computer vision, pages 604–621. Springer, 2022. 1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12

    work page 2022

  40. [40]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652–660,

  41. [41]

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

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 7, 8

    work page 2017

  42. [42]

    Deep hough voting for 3d object detection in point clouds

    Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 9277–9286, 2019. 1

    work page 2019

  43. [43]

    Contrast with Reconstruct: Contrastive 3D Representation Learning Guided by Generative Pretraining

    Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 1, 3, 5, 7, 8, 12

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  44. [44]

    3d object detection for autonomous driving: A survey

    Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 130: 108796, 2022. 1

    work page 2022

  45. [45]

    Explore better relative position embeddings from encoding perspective for transformer models

    Anlin Qu, Jianwei Niu, and Shasha Mo. Explore better relative position embeddings from encoding perspective for transformer models. In Proceedings of the 2021 Confer- ence on Empirical Methods in Natural Language Processing, pages 2989–2997, 2021. 11

    work page 2021

  46. [46]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020. 11

    work page 2020

  47. [47]

    Info3d: Representation learning on 3d ob- jects using mutual information maximization and contrastive learning

    Aditya Sanghi. Info3d: Representation learning on 3d ob- jects using mutual information maximization and contrastive learning. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceed- ings, Part XXIX 16, pages 626–642. Springer, 2020. 1

    work page 2020

  48. [48]

    3d optical machine vision sensors with intelligent data management for robotic swarm navigation improvement

    Oleg Yu Sergiyenko and Vera V Tyrsa. 3d optical machine vision sensors with intelligent data management for robotic swarm navigation improvement. IEEE Sensors Journal, 21 (10):11262–11274, 2020. 1

    work page 2020

  49. [49]

    Self-supervised few-shot learning on point clouds

    Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212–7221, 2020. 5

    work page 2020

  50. [50]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3

    Pith/arXiv arXiv 2014

  51. [51]

    Roformer: Enhanced transformer with rotary position embedding

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063,

  52. [52]

    Proxy graph matching with proximal matching networks

    Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu- Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Confer- ence on Artificial Intelligence (AAAI), 2021. 1

    work page 2021

  53. [53]

    Data pruning via moving- one-sample-out

    Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving- one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  54. [54]

    Semantic diffusion network for semantic segmentation

    Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023

    work page 2023

  55. [55]

    Ensemble quadratic assignment network for graph matching

    Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Com- puter Vision (IJCV), 2024

    work page 2024

  56. [56]

    Saco loss: Sample-wise affinity consistency for vision-language pre-training

    Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xi- aojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    work page 2024

  57. [57]

    Data pruning by information maximization

    Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiao- juan Qi. Data pruning by information maximization. In In- 15 ternational Conference on Learning Representations (ICLR), 2025

    work page 2025

  58. [58]

    Diff-in: Data influence estimation with differential ap- proximation, 2025

    Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential ap- proximation, 2025. 1

    work page 2025

  59. [59]

    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. In Proceedings of the IEEE/CVF international conference on computer vision , pages 1588– 1597, 2019. 5

    work page 2019

  60. [60]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 7, 8

    work page 2017

  61. [62]

    Unsupervised point cloud pre-training via oc- clusion completion

    Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via oc- clusion completion. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 9782–9792,

    work page

  62. [63]

    Dynamic graph cnn for learning on point clouds

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1–12, 2019. 1, 7, 8

    work page 2019

  63. [64]

    Take-a-photo: 3d-to-2d generative pre-training of point cloud models

    Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Ji- wen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pages 5640–5650,

    work page

  64. [65]

    Vertical layering of quantized neural networks for heterogeneous inference

    Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1

    work page 2023

  65. [66]

    Rethinking and improving relative posi- tion encoding for vision transformer

    Kan Wu, Houwen Peng, Minghao Chen, Jianlong Fu, and Hongyang Chao. Rethinking and improving relative posi- tion encoding for vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 10033–10041, 2021. 11

    work page 2021

  66. [67]

    Mixture-of- scores: Robust image-text data quality score via three lines of code

    Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of- scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 1

    work page 2025

  67. [68]

    Point transformer v3: Simpler, faster, stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xi- hui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 1

    Pith/arXiv arXiv 2023

  68. [69]

    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. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015. 5

    work page 1912

  69. [70]

    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 574–591. Springer, 2020. 1, 2

    work page 2020

  70. [71]

    Simmim: A simple framework for masked image modeling

    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. In CVPR, 2022. 2

    work page 2022

  71. [72]

    To- wards robustness and generalization of point cloud represen- tation: A geometry coding method and a large-scale object- level dataset

    Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. To- wards robustness and generalization of point cloud represen- tation: A geometry coding method and a large-scale object- level dataset. Computational Visual Media , 10(1):27–43,

    work page

  72. [73]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding

    Le Xue, Mingfei Gao, Chen Xing, Roberto Mart ´ın-Mart´ın, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179–1189, 2023. 7

    work page 2023

  73. [74]

    Xlnet: Generalized autoregressive pre- training for language understanding

    Zhilin Yang. Xlnet: Generalized autoregressive pre- training for language understanding. arXiv preprint arXiv:1906.08237, 2019. 11

    Pith/arXiv arXiv 1906

  74. [75]

    A scalable active framework for region annotation in 3d shape collections.ACM Transactions on Graphics (ToG), 35(6):1–12, 2016

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

    work page 2016

  75. [76]

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

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313–19322, 2022. 4, 5, 6, 7, 8

    work page 2022

  76. [77]

    Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders

    Yaohua Zha, Huizhen Ji, Jinmin Li, Rongsheng Li, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Towards compact 3d representations via point feature enhancement masked au- toencoders. arXiv preprint arXiv:2312.10726, 2023. 9, 10

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  77. [78]

    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. Advances in neural information processing sys- tems, 35:27061–27074, 2022. 1, 2, 7, 8, 12

    work page 2022

  78. [79]

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

    Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hong- sheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769–21780, 2023. 7, 8, 9, 10, 12

    work page 2023

  79. [80]

    Cr2pq: Continuous relative rotary positional query for dense visual representa- tion learning

    Shaofeng Zhang, Qiang Zhou, Sitong Wu, Haoru Tan, Zhibin Wang, Jinfa Huang, and Junchi Yan. Cr2pq: Continuous relative rotary positional query for dense visual representa- tion learning. In The Thirteenth International Conference on Learning Representations. 1

    work page

  80. [81]

    The diversified ensemble neural network

    Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33:16001–16011, 2020

    work page 2020

Showing first 80 references.