pith. sign in

arxiv: 2605.25127 · v1 · pith:BVUBXF7Ynew · submitted 2026-05-24 · 💻 cs.CV · cs.LG

PQDT: Pseudo-Query Dual Transformer for Robust Point Cloud Restoration

Haoqing Wu , Alexa Nawotki , Jochen Garcke This is my paper

Pith reviewed 2026-06-30 11:50 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords point cloud restorationtransformer network3D completionpoint cloud denoisingdeformation correctionpseudo-query moduleunified 3D restoration
0
0 comments X

The pith

The Pseudo-Query module lets one Transformer network restore point clouds suffering from incompleteness, noise, and deformation at once.

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

This paper introduces a unified network called PQDT that takes degraded point clouds as input and outputs clean, detailed shapes. The key idea is a Pseudo-Query module that splits the process of translating geometry into two cooperative stages inside a Transformer, which helps keep local details and makes the output more robust. A reader would care because real sensor data is often incomplete or noisy, and having one model handle many types of damage could simplify 3D perception pipelines. The method avoids the common problem of losing fine geometry through global features. Tests show it works better than previous approaches on benchmarks with combined degradations.

Core claim

The paper claims that implementing a Pseudo-Query module within a Transformer backbone reformulates geometric translation into two cooperative stages, which enhances structural clarity, robustness, and local detail preservation in point cloud restoration, allowing a single point-only network to handle diverse degradations including completion, deformation, and denoising better than existing methods.

What carries the argument

The Pseudo-Query module that reformulates geometric translation into two cooperative stages.

If this is right

  • It effectively handles complex combinations of completion, deformation, and denoising degradations.
  • Surpasses state-of-the-art performance in general 3D restoration on curated benchmarks.
  • Provides a novel unified, point-only backbone for robust 3D restoration.
  • Enables more versatile 3D perception for downstream applications.

Where Pith is reading between the lines

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

  • The two-stage reformulation could be adapted to other geometric tasks like surface reconstruction from images.
  • If the cooperative stages preserve details well, it might improve performance on very sparse point clouds from LiDAR.
  • Testing on real-world sensor data not seen in training would reveal how well the robustness generalizes.

Load-bearing premise

The two cooperative stages in the Pseudo-Query module improve structural clarity and local details without creating new artifacts or sensitivity to input changes.

What would settle it

Running the model on point clouds with novel combinations of degradations or higher noise levels than in the benchmarks and measuring if it still outperforms other methods or maintains detail without artifacts.

Figures

Figures reproduced from arXiv: 2605.25127 by Alexa Nawotki, Haoqing Wu, Jochen Garcke.

Figure 1
Figure 1. Figure 1: Comparison of point cloud restoration paradigms. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of PQDT. Given an incomplete and noisy input point cloud, local features are extracted via a transition-down module and translated by a dual transformer. Stage I generates pseudo-queries Qps through an encoding–decoding process, providing observation-guided query initialization. Stage II further refines Qps, producing consolidated queries Q to predicted proxies H that are upsampled to generate fin… view at source ↗
Figure 3
Figure 3. Figure 3: Geometric-embedding self-attention block. Geometric attention head uses distance embedding (DE) and angular embed￾ding (AE) from input point coordinates and regrouped as attention keys Kr and values Vr. For example, Pin, Xin of the first self￾attention block in ME1 correspond to P c src and F c src, respectively. Pin = {Pi}M i=1, the dense geometric structure embedding [40, 66] is defined as r_{i,j} = \mat… view at source ↗
Figure 4
Figure 4. Figure 4: Dynamic Query Selection (DQS) module. Given the input coordinates Pin and features Xin, we expand the points with padding points Ppad and zero-padded features Xpad to a fixed size. A Gumbel top-k sampling strategy is then applied, using aggregated scores to select the most representative points as output queries. independently and ignore local geometric relationships. In contrast, coarse-to-fine upsampling… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of datasets used for evaluation. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative evaluation on ShapeNet-Deform and ShapeNetCar-Occ. Red boxes indicate occluding objects [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative evaluation on PFS. The color represent the Point-to-Surface distance (P2S) between generated point clouds and ground truth mesh, the maximum (red) is clamped with 2% of the radius of the object’s bounding sphere. can effectively infer missing geometry and remains robust against local occlusion noise and global Gaussian noise. These results further validate the effectiveness of our query selecti… view at source ↗
Figure 8
Figure 8. Figure 8: Attention maps from selected queries (red dot) visu￾alized on key points (bottom row). Stage I attention maps show coarse pseudo-query exploration over the encoded latent structure. Stage II maps demonstrate localized, coherent attention aligned with the underlying surface, indicating effective query refinement. The color reflects the attention score after the Softmax operation in the last block of the dec… view at source ↗
Figure 9
Figure 9. Figure 9: Illustration of PFS usage in an industrial application. The [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Training loss under different initializations of query [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative evaluation on ShapeNet-Deform [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative evaluation on ShapeNetCar-Occ. Red boxes indicate occluding objects. [PITH_FULL_IMAGE:figures/full_fig_p017_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative evaluation on PFS. Colors indicate the point [PITH_FULL_IMAGE:figures/full_fig_p018_13.png] view at source ↗
read the original abstract

Point clouds are a fundamental 3D representation in computer vision, enabling a wide range of perception tasks. However, real-world point clouds often suffer from degradations such as incompleteness, noise, outliers, and irregular density, caused by sensor limitations or occlusions. Recovering clean and detailed shapes from such degraded data is crucial for downstream applications. While existing learning-based methods achieve progress on individual tasks like completion or denoising, they typically rely on global bottleneck features, which lose fine-grained geometry and remain sensitive to varying input quality. We propose a unified 3D restoration network that directly takes point clouds as input and adaptively reconstructs high-quality geometry under diverse degradation scenarios. At the core of our approach is a Pseudo-Query module, implemented within a Transformer backbone, which reformulates geometric translation into two cooperative stages to enhance structural clarity, robustness, and local detail preservation. Extensive experiments on curated benchmarks demonstrate that our approach surpasses state-of-the-art performance in general 3D restoration. It effectively handles complex combinations of completion, deformation, and denoising degradations. With this work, we provide a novel unified, point-only backbone for robust 3D restoration, enabling more versatile 3D perception.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes PQDT, a unified point-cloud restoration network built around a Pseudo-Query Dual Transformer. A Pseudo-Query module inside the Transformer backbone reformulates geometric translation into two cooperative stages intended to improve structural clarity, robustness, and local-detail preservation. The work claims to surpass prior state-of-the-art methods on curated benchmarks for general 3D restoration and to handle complex combinations of completion, deformation, and denoising degradations within a single point-only backbone.

Significance. A validated unified backbone that demonstrably copes with simultaneous multi-degradation inputs would constitute a meaningful advance over task-specific pipelines that rely on global bottleneck features. The absence of any quantitative results, baseline comparisons, ablation tables, or benchmark descriptions in the supplied text, however, precludes any assessment of whether that advance has been achieved.

major comments (1)
  1. [Abstract] Abstract: the headline claim that the method 'effectively handles complex combinations of completion, deformation, and denoising degradations' is load-bearing for the central contribution, yet the text supplies no experimental protocol, per-combination metrics, or description of whether the curated benchmarks contain simultaneous multi-degradation inputs versus single or sequentially applied degradations. Without this information the generalization asserted for the Pseudo-Query module cannot be evaluated.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive comment on the abstract. We address the concern point-by-point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim that the method 'effectively handles complex combinations of completion, deformation, and denoising degradations' is load-bearing for the central contribution, yet the text supplies no experimental protocol, per-combination metrics, or description of whether the curated benchmarks contain simultaneous multi-degradation inputs versus single or sequentially applied degradations. Without this information the generalization asserted for the Pseudo-Query module cannot be evaluated.

    Authors: We acknowledge that the abstract is brief and does not detail the experimental protocol. The full manuscript (Section 4) describes the curated benchmarks, which are generated by applying completion, deformation, and denoising degradations simultaneously to each input point cloud (rather than sequentially or in isolation). Quantitative results, including per-combination metrics on these multi-degradation cases, are reported in Tables 1–3 with comparisons to prior methods. To address the referee's concern, we will revise the abstract to explicitly note that the benchmarks feature simultaneous multi-degradation inputs and to reference the experimental section for the evaluation protocol and metrics. revision: yes

Circularity Check

0 steps flagged

No circularity: architecture and claims rest on external benchmarks, not self-referential definitions or fits.

full rationale

The paper proposes a Transformer-based network with a Pseudo-Query module for point cloud restoration and supports its performance claims solely via experiments on curated benchmarks. No equations, derivations, or first-principles results are present that could reduce to inputs by construction. The design choices are presented as empirical engineering decisions rather than mathematically forced outcomes, and no self-citation chains or fitted parameters renamed as predictions appear in the abstract or described structure. This is the standard case of a self-contained empirical ML contribution whose validity is tested externally.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no information on free parameters, axioms, or invented entities used in the method.

pith-pipeline@v0.9.1-grok · 5746 in / 1006 out tokens · 26090 ms · 2026-06-30T11:50:22.681411+00:00 · methodology

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

77 extracted references · 2 canonical work pages · 2 internal anchors

  1. [1]

    Estimating or Propagating Gradients Through Stochastic Neurons for Conditional Computation

    Yoshua Bengio, Nicholas L´eonard, and Aaron Courville. Es- timating or propagating gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    State of the art in surface reconstruction from point clouds.Euro- graphics 2014-State of the Art Reports, 2014

    Matthew Berger, Andrea Tagliasacchi, Lee Seversky, Pierre Alliez, Joshua Levine, Andrei Sharf, and Cl´audio Silva. State of the art in surface reconstruction from point clouds.Euro- graphics 2014-State of the Art Reports, 2014. 2

    2014

  3. [3]

    End- to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End- to-end object detection with transformers. InECCV, pages 213–229, 2020. 3, 2

    2020

  4. [4]

    Anchorformer: Point cloud completion from discriminative nodes

    Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wen- gang Zhou, Jiebo Luo, and Tao Mei. Anchorformer: Point cloud completion from discriminative nodes. InCVPR, pages 13581–13590, 2023. 2, 4, 6, 7, 3, 9

    2023

  5. [5]

    Shape completion using 3d-encoder-predictor cnns and shape synthesis

    Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. InCVPR, pages 5868–5877, 2017. 2

    2017

  6. [6]

    Iterativepfn: True iterative point cloud filtering

    Dasith de Silva Edirimuni, Xuequan Lu, Zhiwen Shao, Gang Li, Antonio Robles-Kelly, and Ying He. Iterativepfn: True iterative point cloud filtering. InCVPR, pages 13530–13539,

  7. [7]

    Superpc: a single diffusion model for point cloud completion, upsampling, de- noising, and colorization

    Yi Du, Zhipeng Zhao, Shaoshu Su, Sharath Golluri, Haoze Zheng, Runmao Yao, and Chen Wang. Superpc: a single diffusion model for point cloud completion, upsampling, de- noising, and colorization. InCVPR, pages 16953–16964,

  8. [8]

    T-corresnet: template guided 3d point cloud completion with correspondence pool- ing query generation strategy

    Fan Duan, Jiahao Yu, and Li Chen. T-corresnet: template guided 3d point cloud completion with correspondence pool- ing query generation strategy. InECCV, pages 90–106, 2024. 2, 6, 7

    2024

  9. [9]

    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. InCVPR, pages 605–613, 2017. 5, 2

    2017

  10. [10]

    Robust moving least-squares fitting with sharp features.ACM TOG, 24(3):544–552, 2005

    Shachar Fleishman, Daniel Cohen-Or, and Cl´audio T Silva. Robust moving least-squares fitting with sharp features.ACM TOG, 24(3):544–552, 2005. 2

    2005

  11. [11]

    Me-pcn: Point completion conditioned on mask emptiness

    Bingchen Gong, Yinyu Nie, Yiqun Lin, Xiaoguang Han, and Yizhou Yu. Me-pcn: Point completion conditioned on mask emptiness. InICCV, pages 12488–12497, 2021. 2

    2021

  12. [12]

    3d-coded: 3d correspondences by deep deformation

    Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. 3d-coded: 3d correspondences by deep deformation. InECCV, pages 230–246, 2018. 3

    2018

  13. [13]

    Body in white (biw)

    SKH Group. Body in white (biw). https : / / www.skhgroup.co.in/our-products/body-in- white-biw. Accessed: Nov. 2025. 4

    2025

  14. [14]

    High-resolution shape completion using deep neural networks for global structure and local geometry inference

    Xiaoguang Han, Zhen Li, Haibin Huang, Evangelos Kaloger- akis, and Yizhou Yu. High-resolution shape completion using deep neural networks for global structure and local geometry inference. InICCV, pages 85–93, 2017. 2

    2017

  15. [15]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. InNeurIPS, pages 6840–6851,

  16. [16]

    Render4completion: Synthesizing multi-view depth maps for 3d shape completion

    Tao Hu, Zhizhong Han, Abhinav Shrivastava, and Matthias Zwicker. Render4completion: Synthesizing multi-view depth maps for 3d shape completion. InICCVW, 2019. 2

    2019

  17. [17]

    Local frequency interpretation and non-local self-similarity on graph for point cloud inpainting.IEEE TIP, 28(8):4087–4100, 2019

    Wei Hu, Zeqing Fu, and Zongming Guo. Local frequency interpretation and non-local self-similarity on graph for point cloud inpainting.IEEE TIP, 28(8):4087–4100, 2019. 2

    2019

  18. [18]

    Fea- ture graph learning for 3d point cloud denoising.IEEE TSP, 68:2841–2856, 2020

    Wei Hu, Xiang Gao, Gene Cheung, and Zongming Guo. Fea- ture graph learning for 3d point cloud denoising.IEEE TSP, 68:2841–2856, 2020. 2

    2020

  19. [19]

    Rfnet: Recurrent forward network for dense point cloud completion

    Tianxin Huang, Hao Zou, Jinhao Cui, Xuemeng Yang, Meng- meng Wang, Xiangrui Zhao, Jiangning Zhang, Yi Yuan, Yifan Xu, and Yong Liu. Rfnet: Recurrent forward network for dense point cloud completion. InICCV, pages 12508–12517,

  20. [20]

    Pf-net: Point fractal network for 3d point cloud completion

    Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. InCVPR, pages 7662–7670, 2020. 1, 2

    2020

  21. [21]

    Point cloud completion with pretrained text-to-image diffusion models

    Yoni Kasten, Ohad Rahamim, and Gal Chechik. Point cloud completion with pretrained text-to-image diffusion models. InNeurIPS, pages 12171–12191, 2023. 2

    2023

  22. [22]

    Learning part-based templates from large collections of 3d shapes

    Vladimir G Kim, Wilmot Li, Niloy J Mitra, Siddhartha Chaud- huri, Stephen DiVerdi, and Thomas Funkhouser. Learning part-based templates from large collections of 3d shapes. ACM TOG, 32(4):1–12, 2013. 2

    2013

  23. [23]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InICLR, 2015. 1

    2015

  24. [24]

    Robo3d: Towards robust and reliable 3d perception against corruptions

    Lingdong Kong, Youquan Liu, Xin Li, Runnan Chen, Wenwei Zhang, Jiawei Ren, Liang Pan, Kai Chen, and Ziwei Liu. Robo3d: Towards robust and reliable 3d perception against corruptions. InICCV, pages 19994–20006, 2023. 2

    2023

  25. [25]

    Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without replacement

    Wouter Kool, Herke Van Hoof, and Max Welling. Stochastic beams and where to find them: The gumbel-top-k trick for sampling sequences without replacement. InICML, pages 3499–3508, 2019. 5

    2019

  26. [26]

    Procedural noise using sparse gabor convolution.ACM SIGGRAPH, 28(3):54–64, 2009

    Ares Lagae, Sylvain Lefebvre, George Drettakis, and Philip Dutr´e. Procedural noise using sparse gabor convolution.ACM SIGGRAPH, 28(3):54–64, 2009. 6, 3 6 Table 10. Results of our method and state-of-the-art methods on ShapeNet-55. We report the detailed results for each method on 10 categories and the overall results on 55 categories for three difficulty...

    2009

  27. [27]

    M. Levoy. Display of surfaces from volume data.IEEE Computer Graphics and Applications, 8(3):29–37, 1988. 6

    1988

  28. [28]

    Proxyformer: Proxy alignment assisted point cloud comple- tion with missing part sensitive transformer

    Shanshan Li, Pan Gao, Xiaoyang Tan, and Mingqiang Wei. Proxyformer: Proxy alignment assisted point cloud comple- tion with missing part sensitive transformer. InCVPR, pages 9466–9475, 2023. 2, 4, 6, 7

    2023

  29. [29]

    Database-assisted object retrieval for real-time 3d reconstruction

    Yangyan Li, Angela Dai, Leonidas Guibas, and Matthias Nießner. Database-assisted object retrieval for real-time 3d reconstruction. InCGF, pages 435–446, 2015. 2

    2015

  30. [30]

    SGDR: stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. InICLR, 2017. 1

    2017

  31. [31]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. InCVPR, pages 2837–2845, 2021. 3

    2021

  32. [32]

    Pd-flow: A point cloud denoising framework with normalizing flows

    Aihua Mao, Zihui Du, Yu-Hui Wen, Jun Xuan, and Yong- Jin Liu. Pd-flow: A point cloud denoising framework with normalizing flows. InECCV, pages 398–415, 2022. 2

    2022

  33. [33]

    Denoising point clouds in latent space via graph convolution and in- vertible neural network

    Aihua Mao, Biao Yan, Zijing Ma, and Ying He. Denoising point clouds in latent space via graph convolution and in- vertible neural network. InCVPR, pages 5768–5777, 2024. 2

    2024

  34. [34]

    Symmetry in 3d geometry: Extraction and applications.CGF, 32(6), 2013

    Niloy Mitra, Mark Pauly, Michael Wand, and Duygu Ceylan. Symmetry in 3d geometry: Extraction and applications.CGF, 32(6), 2013. 2

    2013

  35. [35]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 1

    2019

  36. [36]

    Categorical results of our method on ShapeNet-55 and ShapeNet-Deform

    Mark Pauly, Niloy J Mitra, Joachim Giesen, Markus H Gross, 7 Table 11. Categorical results of our method on ShapeNet-55 and ShapeNet-Deform. We report the detailed results under CD ℓ2 and F-Score@1% metric for Simple, Moderate and Hard settings. Category ShapeNet-55 ShapeNet-Deform Simple Moderate Hard Simple Moderate Hard CDℓ2 (↓) F1 (↑) CD ℓ2 (↓) F1 (↑)...

    2005

  37. [37]

    Learning graph-convolutional representations Table 13

    Francesca Pistilli, Giulia Fracastoro, Diego Valsesia, and Enrico Magli. Learning graph-convolutional representations Table 13. Results of our method and state-of-the-art methods on ShapeNetCar-Occ. We report the detailed results for each method on Car category for three difficulty degrees. We use CD-S, CD-M and CD-H to represent the CD results under the ...

    2020

  38. [38]

    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. InCVPR, pages 652–660, 2017. 1, 2

    2017

  39. [39]

    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. InNeurIPS, 2017. 1, 2, 4

    2017

  40. [40]

    Geotransformer: Fast and robust point cloud registration with geometric trans- former.IEEE TPAMI, 45(8):9806–9821, 2023

    Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, Slobodan Ilic, Dewen Hu, and Kai Xu. Geotransformer: Fast and robust point cloud registration with geometric trans- former.IEEE TPAMI, 45(8):9806–9821, 2023. 4, 1

    2023

  41. [41]

    Pointcleannet: Learning to denoise and remove outliers from dense point clouds

    Marie-Julie Rakotosaona, Vittorio La Barbera, Paul Guerrero, Niloy J Mitra, and Maks Ovsjanikov. Pointcleannet: Learning to denoise and remove outliers from dense point clouds. In CGF, pages 185–203, 2020. 2

    2020

  42. [42]

    Kripasindhu Sarkar, Kiran Varanasi, and Didier Stricker. 9 (a) (b) (c) (d) InputOutput AdaPoinTr P2S AdaPoinTr Output AnchorFormer P2S AnchorFormer Output PQDT P2S PQDT Ground Truth Output SnowflakeNet P2S SnowflakeNet Output SeedFormer P2S SeedFormer 0% 2% Figure 13. Qualitative evaluation on PFS. Colors indicate the point- to-surface (P2S) distance betw...

    2017

  43. [43]

    Learning 3d shape comple- tion from laser scan data with weak supervision

    David Stutz and Andreas Geiger. Learning 3d shape comple- tion from laser scan data with weak supervision. InCVPR, pages 1955–1964, 2018. 2

    1955

  44. [44]

    Data-driven structural priors for shape completion

    Minhyuk Sung, Vladimir G Kim, Roland Angst, and Leonidas Guibas. Data-driven structural priors for shape completion. ACM TOG, 34(6):1–11, 2015. 2

    2015

  45. [45]

    What do single-view 3d reconstruction networks learn? InCVPR, pages 3405– 3414, 2019

    Maxim Tatarchenko, Stephan R Richter, Ren´e Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3d reconstruction networks learn? InCVPR, pages 3405– 3414, 2019. 6, 2

    2019

  46. [46]

    Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese

    Lyne P. Tchapmi, Vineet Kosaraju, Hamid Rezatofighi, Ian Reid, and Silvio Savarese. Topnet: Structural point cloud decoder. InCVPR, pages 383–392, 2019. 2

    2019

  47. [47]

    Deformation-aware 3d model embedding and retrieval

    Mikaela Angelina Uy, Jingwei Huang, Minhyuk Sung, Tolga Birdal, and Leonidas Guibas. Deformation-aware 3d model embedding and retrieval. InECCV, pages 397–413, 2020. 3

    2020

  48. [48]

    Joint learning of 3d shape retrieval and deforma- tion

    Mikaela Angelina Uy, Vladimir G Kim, Minhyuk Sung, Noam Aigerman, Siddhartha Chaudhuri, and Leonidas J Guibas. Joint learning of 3d shape retrieval and deforma- tion. InCVPR, pages 11713–11722, 2021. 3

    2021

  49. [49]

    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. InNeurIPS, page 6000–6010, 2017. 1

    2017

  50. [50]

    3dn: 3d deformation network

    Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. 3dn: 3d deformation network. InCVPR, pages 1038–1046, 2019. 3

    2019

  51. [51]

    Cas- caded refinement network for point cloud completion

    Xiaogang Wang, Marcelo H Ang Jr, and Gim Hee Lee. Cas- caded refinement network for point cloud completion. In CVPR, pages 790–799, 2020. 2

    2020

  52. [52]

    Dynamic graph cnn for learning on point clouds.ACM TOG, 38(5): 1–12, 2019

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds.ACM TOG, 38(5): 1–12, 2019. 4, 1

    2019

  53. [53]

    Pcdreamer: Point cloud completion through multi-view diffusion priors

    Guangshun Wei, Yuan Feng, Long Ma, Chen Wang, Yuanfeng Zhou, and Changjian Li. Pcdreamer: Point cloud completion through multi-view diffusion priors. InCVPR, pages 27243– 27253, 2025. 5, 7

    2025

  54. [54]

    Point cloud completion by skip-attention network with hi- erarchical folding

    Xin Wen, Tianyang Li, Zhizhong Han, and Yu-Shen Liu. Point cloud completion by skip-attention network with hi- erarchical folding. InCVPR, pages 1939–1948, 2020. 1, 2

    1939

  55. [55]

    Simple statistical gradient-following al- gorithms for connectionist reinforcement learning.Machine learning, 8(3):229–256, 1992

    Ronald J Williams. Simple statistical gradient-following al- gorithms for connectionist reinforcement learning.Machine learning, 8(3):229–256, 1992. 2

    1992

  56. [56]

    Lever- aging single-view images for unsupervised 3d point cloud completion.IEEE TMM, 27:940–953, 2023

    Lintai Wu, Qijian Zhang, Junhui Hou, and Yong Xu. Lever- aging single-view images for unsupervised 3d point cloud completion.IEEE TMM, 27:940–953, 2023. 2

    2023

  57. [57]

    Point transformer v2: Grouped vector attention and partition-based pooling

    Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Heng- shuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. InNeurIPS, 2022. 1

    2022

  58. [58]

    Point transformer v3: Simpler, faster, stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 1 10

    2024

  59. [59]

    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. InCVPR, pages 1912–1920, 2015. 5

    1912

  60. [60]

    Snowflakenet: Point cloud completion by snowflake point deconvolution with skip- transformer

    Peng Xiang, Xin Wen, Yu-Shen Liu, Yan-Pei Cao, Pengfei Wan, Wen Zheng, and Zhizhong Han. Snowflakenet: Point cloud completion by snowflake point deconvolution with skip- transformer. InICCV, pages 5499–5509, 2021. 1, 2, 5, 6, 7, 9

    2021

  61. [61]

    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. InECCV, pages 365–381, 2020. 2, 6, 4, 7

    2020

  62. [62]

    Reparameterizable subset sampling via continuous relaxations

    Sang Michael Xie and Stefano Ermon. Reparameterizable subset sampling via continuous relaxations. InIJCAI, page 3919–3925, 2019. 5

    2019

  63. [63]

    Explicitly guided information interaction network for cross-modal point cloud completion

    Hang Xu, Chen Long, Wenxiao Zhang, Yuan Liu, Zhen Cao, Zhen Dong, and Bisheng Yang. Explicitly guided information interaction network for cross-modal point cloud completion. InECCV, pages 414–432, 2024. 2

    2024

  64. [64]

    3d object reconstruction from a single depth view with adversarial learning

    Bo Yang, Hongkai Wen, Sen Wang, Ronald Clark, Andrew Markham, and Niki Trigoni. 3d object reconstruction from a single depth view with adversarial learning. InICCVW, pages 679–688, 2017. 2

    2017

  65. [65]

    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. InCVPR, pages 206–215, 2018. 5, 6, 4, 7

    2018

  66. [66]

    Rotation-invariant trans- former for point cloud matching

    Hao Yu, Zheng Qin, Ji Hou, Mahdi Saleh, Dongsheng Li, Benjamin Busam, and Slobodan Ilic. Rotation-invariant trans- former for point cloud matching. InCVPR, pages 5384–5393,

  67. [67]

    Pointr: Diverse point cloud completion with geometry-aware transformers

    Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. InCVPR, pages 12498–12507,

  68. [68]

    Adapointr: Diverse point cloud completion with adap- tive geometry-aware transformers.IEEE TPAMI, 45(12): 14114–14130, 2023

    Xumin Yu, Yongming Rao, Ziyi Wang, Jiwen Lu, and Jie Zhou. Adapointr: Diverse point cloud completion with adap- tive geometry-aware transformers.IEEE TPAMI, 45(12): 14114–14130, 2023. 2, 3, 5, 6, 7, 4, 9

    2023

  69. [69]

    Pcn: Point completion network

    Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In3DV, pages 728–737, 2018. 1, 2, 5, 6, 7, 4, 9

    2018

  70. [70]

    Point cloud denoising with principal component analysis and a novel bilateral filter.Traitement du signal, 36:393–398, 2019

    Feng Zhang, Chao Zhang, Huamin Yang, and Lin Zhao. Point cloud denoising with principal component analysis and a novel bilateral filter.Traitement du signal, 36:393–398, 2019. 2

    2019

  71. [71]

    View-guided point cloud completion

    Xuancheng Zhang, Yutong Feng, Siqi Li, Changqing Zou, Hai Wan, Xibin Zhao, Yandong Guo, and Yue Gao. View-guided point cloud completion. InCVPR, pages 15890–15899, 2021. 2

    2021

  72. [72]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. InICCV, pages 16259– 16268, 2021. 1, 2, 4

    2021

  73. [73]

    Rolling normal filtering for point clouds

    Yinglong Zheng, Guiqing Li, Xuemiao Xu, Shihao Wu, and Yongwei Nie. Rolling normal filtering for point clouds. CAGD, 62(C):16–28, 2018. 2

    2018

  74. [74]

    Position-aware guided point cloud completion with clip model

    Feng Zhou, Qi Zhang, Ju Dai, Lei Li, Qing Fan, and Junliang Xing. Position-aware guided point cloud completion with clip model. InAAAI, pages 10734–10742, 2025. 2

    2025

  75. [75]

    Seedformer: Patch seeds based point cloud completion with upsample transformer

    Haoran Zhou, Yun Cao, Wenqing Chu, Junwei Zhu, Tong Lu, Ying Tai, and Chengjie Wang. Seedformer: Patch seeds based point cloud completion with upsample transformer. InECCV, pages 416–432, 2022. 2, 4, 5, 6, 7, 3, 9

    2022

  76. [76]

    Open3D: A Modern Library for 3D Data Processing

    Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing.arXiv preprint arXiv:1801.09847, 2018. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  77. [77]

    Svdformer: Complementing point cloud via self-view augmentation and self-structure dual-generator

    Zhe Zhu, Honghua Chen, Xing He, Weiming Wang, Jing Qin, and Mingqiang Wei. Svdformer: Complementing point cloud via self-view augmentation and self-structure dual-generator. InICCV, pages 14508–14518, 2023. 2 11

    2023