Pith. sign in

REVIEW 5 major objections 5 minor 56 references

Denoising-While-Completing Network (DWCNet): Robust Point Cloud Completion Under Corruption

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that point cloud completion becomes robust to real-world corruption when the network explicitly separates clean from noisy features before decoding, and backs the claim with a new corrupted benchmark and a network that…

desk verdict CPCCD is a genuine new benchmark for point cloud completion robustness, but DWCNet's own Table 3 shows catastrophic pre-fine-tuning failures that invalidate the paper's central robustness claim until explained. read the letter →

arxiv 2507.16743 v1 pith:CWKF5XOC submitted 2025-07-22 cs.CV

Keneni W. Tesema , Lyndon Hill , Mark W. Jones , Gary K.L. Tam This is my paper
classification cs.CV
keywords pointcloudcompletiondenoisingrobustnessbenchmarkCPCCDdatasetDWCNetcontrastivelearningself-attentioncorruption
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

Point cloud completion networks are trained on clean synthetic partials and therefore fail when real scans arrive with background points, occlusions, jitter, and scale mismatch. This paper argues that completion should include denoising in the same forward pass, and backs the argument with a corrupted benchmark plus a network. The benchmark, CPCCD, applies eight corruption types, derived from visual inspection of a real indoor dataset, to the partials of the standard PCN completion dataset, with randomized parameter ranges. The network, DWCNet, inserts a Noise Management Module into a transformer-based completion backbone so that encoded features are classified as clean or noisy and only clean features drive the decoder. If the claims hold, completion systems built this way would work on realistic indoor scans rather than only on sanitized synthetic data.

What carries the argument

The Noise Management Module (NMM) is the load-bearing object: a dual-path feature separator inserted between a transformer encoder and decoder. Its clean path applies multi-head self-attention followed by a position-wise feed-forward network with residual connections and layer normalization, while its noisy path applies parallel one-dimensional convolutions with kernel sizes 1, 3, and 5 and merges them with learnable weights. A contrastive loss, combining a positive term that maximizes cosine similarity between clean features and clean ground-truth features and a negative term that minimizes similarity between clean and noisy features, trains the module to classify encoded features. Only the clean features are passed to the decoder, so denoising and completion happen in one pass rather than as separate stages.

What would settle it

Run DWCNet on raw real-world indoor scans that contain corruption types outside the CPCCD training set, such as motion blur from a moving sensor or missing columns from depth-camera dropouts, and compare its output against a registered CAD model. If its Chamfer Distance advantage over a fine-tuned AdaPoinTr vanishes or its completed shapes trail the baseline, the claim that separating clean from noisy features generalizes beyond the benchmark's corruption distribution is falsified. A simpler check is to perturb CPCCD parameter ranges and see whether performance degrades continuously or collapses at the boundary.

Watch

Extended reading notes

Core claim

The paper's central claim is that a completion network can be made robust to multiple simultaneous corruptions by explicitly separating clean from noisy features before decoding, and that this beats both training-only robustness and the base completion architecture. DWCNet adopts AdaPoinTr as its backbone and adds the Noise Management Module (NMM), a dual-path module whose clean path uses multi-head self-attention and a feed-forward network to model structural relationships, and whose noisy path uses multi-scale one-dimensional convolutions to capture noise at different scales. A contrastive loss pulls clean features toward ground-truth features while pushing clean and noisy features apart, with temperature scaling to balance the two terms. On the new CPCCD benchmark, DWCNet reports the best Chamfer Distance, F-score, and fidelity among the evaluated networks after fine-tuning, and it also improves over its backbone on the clean PCN dataset. These results support the paper's further claim that CPCCD is a usable first robustness benchmark for point cloud completion.

Load-bearing premise

The benchmark's value depends on the assumption that the eight corruption types and parameter ranges, chosen by visual inspection of one real-world indoor dataset, faithfully represent the corruptions that completion systems will meet in real indoor scans.

Editorial extensions

If this is right

  • Owners of existing transformer-based completion networks can insert NMM-style feature separation and gain robustness on corrupted inputs without changing the completion loss.
  • Fine-tuning any evaluated network on CPCCD improves its resilience, but the paper's numbers indicate that a dedicated noise-management module yields lower Chamfer Distance than fine-tuning alone on the hardest combined-corruption setting.
  • CPCCD gives the completion community a common test bed with eight corruption types and randomized severity, so robustness across methods can be compared instead of inferred from clean-data results.
  • Because DWCNet is built on AdaPoinTr, the design principle transfers: robustness can come from managing noisy features inside the encoder, not only from training data augmentation.

Reading between the lines

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

  • The corruption-generation recipe in CPCCD could be reused as a data-augmentation strategy for other point cloud tasks, such as classification and segmentation, where a clean/noisy feature separation module may be equally applicable; the paper does not test this.
  • The real-world claim rests on qualitative results only; a quantitative evaluation on ScanNet or KITTI, using a registered CAD ground truth, would be the natural next test and is not in the paper.
  • Because the corruption parameters were chosen by visual inspection of one indoor dataset, the authors' broader claim about real-world noise would be strengthened by calibrating the ranges to sensor-specific measurement noise, such as depth-camera jitter, and checking that performance transfers.
  • The ablation results suggest the clean path carries most of the benefit while the noisy path adds little; a simpler module with only the contrastive positive loss might achieve most of the gain, though the authors argue the negative loss prevents representation collapse.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The manuscript introduces CPCCD, a corrupted version of the PCN partial point clouds with eight corruption types drawn from a visual inspection of ScanObjectNN, and proposes DWCNet, a completion network whose Noise Management Module (NMM) separates clean and noisy features via contrastive learning and self-attention. The authors evaluate six completion networks on CPCCD before and after fine-tuning, report that DWCNet achieves state-of-the-art results on clean and corrupted synthetic data, and provide qualitative results on ScanObjectNN. The paper presents CPCCD as the first robustness benchmark for point cloud completion and DWCNet as a method that completes while denoising.

Significance. If the empirical claims held, the paper would make two useful contributions: a systematic robustness benchmark for point cloud completion and a concrete architectural mechanism for improving robustness to multiple simultaneous corruptions. The systematic evaluation of six architectures across eight corruption types is valuable, and the idea of separating clean and noisy features inside a completion transformer is reasonable and worth pursuing. The paper does not ship code or data at review time, so none of the claims are independently verifiable. More importantly, several load-bearing experimental numbers are internally inconsistent, and until those are resolved the benchmark and the method cannot be reliably assessed.

major comments (5)
  1. [Table 3 and Table A.15] DWCNet before fine-tuning collapses on isometric scaling (IS) and random combined corruption (RCC): Table 3 reports CD-L1 of 273.589 for IS and 57.919 for RCC, and Table 5 reports fidelity of 31652.809 for IS and 485.849 for RCC, while the base model AdaPoinTr reports 23.016/36.682 CD-L1 and 1.344/5.859 fidelity; Table A.15 shows the same failure in CD-L2 (31673.3 for IS). Since DWCNet matches AdaPoinTr on clean data (6.493 vs 6.516), the failure is specific to scale and combined corruption and is never explained in Section 6.1. If these numbers are correct, DWCNet is catastrophically non-robust to two of the eight benchmark corruptions before fine-tuning, contradicting the unqualified robustness claim; if they are a pipeline artifact, the same artifact may contaminate the fine-tuned SOTA numbers in Tables 6-8. This anomaly must be resolved before the central claim can be assessed.
  2. [Section 7, Table 9] The ablation baseline labeled 'DWCNET w/o NMM (equivalent to AdaPoinTr)' reports CD-L1 of 8.351 on PCN and 13.403 on Rcc, whereas the official AdaPoinTr reports 6.516 on clean PCN in Table 3 and 10.317 on RCC after fine-tuning in Table 6. The w/o NMM baseline is therefore not equivalent to the official AdaPoinTr, and the claimed benefit of NMM (10.263 vs 13.403) is measured against a model that is substantially worse than the model it is supposed to reproduce. The authors should report AdaPoinTr trained under the same protocol, or otherwise explain the discrepancy.
  3. [Section 7, Tables 13-14] The temperature parameter t is selected after evaluating completion results on the CPCCD Rcc category and then across all CPCCD categories at 350 epochs. If those evaluations use the CPCCD test set, then the fine-tuned numbers in Tables 6-8 are produced by a model whose hyperparameter was chosen on the test set, which inflates the reported advantage and weakens the comparison with baselines that use their default hyperparameters. Please clarify which split was used for this selection and, if necessary, move the selection to a validation set.
  4. [Section 6.3] The real-world evaluation on ScanObjectNN is purely qualitative: experiments are reported on 20 chairs, 20 tables, 20 desks, and 20 sofas, with only example images in Figure 10. This does not support the abstract and conclusion claims that DWCNet achieves state-of-the-art results on real-world datasets. A quantitative protocol for completion without ground truth, or a clearly stated limitation, is needed before this claim can be accepted.
  5. [Section 3, Table 2] The CPCCD corruption types and parameter ranges are selected by visually inspecting ScanObjectNN and then applied to PCN partials, but no quantitative evidence is provided that the simulated corruptions match the distribution of real indoor scans. If they do not, then robustness on CPCCD may not transfer to real-world data, so the benchmark's validity as a robustness benchmark and the real-world generalization claim are not yet established.
minor comments (5)
  1. [Tables 3-8] The metric header 'f idelity' contains a spurious space; please fix to 'Fidelity'.
  2. [Section 4.1, Eq. (6)] The negative contrastive loss as written omits an explicit normalization by the number of negative pairs and does not clearly define the index set over which the sum runs; please clarify the notation.
  3. [Abstract and Introduction] The introduction says the benchmark enables 'objective and qualitative robustness evaluation'; the intended word is presumably 'quantitative'.
  4. [Throughout] The model name appears as 'DWCNet', 'DWCNET', and 'DWCNET(Ours)' in different places; please standardize the spelling.
  5. [Appendix D] The GRNet coarse/dense misalignment discussion is only qualitative and is not connected to the corruption types; consider stating this explicitly as a limitation.

Circularity Check

1 steps flagged · score 1.0 of 10

Core derivation is self-contained; only minor circularity is temperature selection on the CPCCD evaluation split.

  1. fitted input called prediction [Section 7.3 (Contrastive Loss Variants), Tables 13 and 14; also Section 4.1: 'we use a temperature scaling t = 1 following findings from our ablation study']
    "We found that t = 1.0 and t = 100.0 yielded the most favorable results on the Rcc category, with chamfer distances of 9.754 and 9.752, respectively. While t = 100.0 offered a marginal advantage on Rcc, the model trained at t = 1.0 outperformed across all other CPCCD categories and was therefore selected as the optimal configuration as shown in Table 14."

    The contrastive temperature t is selected by evaluating candidate values on the CPCCD Rcc test split (Table 13), and the same split is then used to report DWCNet's state-of-the-art results (Tables 6-8 and 14). The reported Rcc performance for the chosen configuration is therefore in-sample with respect to hyperparameter selection rather than an independent prediction. This is a mild form of fitted-input-called-prediction, but its impact is small: across t values at 350 epochs Rcc CD ranges only from about 9.752 to 10.328, and the difference between t=1 and t=100 is 0.002. The architecture and training procedure remain independently specified, so this is not a load-bearing circularity.

full rationale

The paper's claimed derivation chain is largely self-contained. DWCNet's denoising-while-completing behavior is obtained by appending a Noise Management Module (NMM) to AdaPoinTr and training end-to-end with a completion loss plus a contrastive loss whose positive anchor is the clean partial ground truth; this is ordinary supervised training, not a definitional identity with the output. CPCCD is constructed by hand-designed corruption types applied to PCN partials, and the state-of-the-art comparison after fine-tuning is an empirical comparison rather than a derivation from the benchmark's definition. The only identifiable circular step is the selection of the temperature t on the CPCCD evaluation split before reporting the same split's results; this is a test-set-selection issue of small magnitude. Self-citation is confined to the authors' survey [2] cited in the introductory framing and is not load-bearing. The unexplained catastrophic IS/RCC values for DWCNet before fine-tuning (Table 3 and A.15) are an internal-consistency or correctness concern, not a circularity, and do not make the derivation circular.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim depends on the hand-selected corruption taxonomy in CPCCD, which is not quantitatively validated against real data, and on the standard completion metrics. No new physical entities are introduced.

free parameters (3)
  • NMM temperature t = 1.0
    Selected via ablation on the CPCCD evaluation categories (Section 7.3, Table 14); a form of test-set tuning.
  • Contrastive loss weight = 1.0 (implicit)
    L_total = L_completion + L_NMM without a tuned weight; a design choice that affects the balance.
  • CPCCD corruption parameter ranges = Ranges in Table 2 (e.g., Np=[1/16,1/12,1/8,1/4], Is=[0.25,2])
    Hand-selected from visual inspection; not derived from real data distributions.
assumptions (3)
  • domain assumption The eight corruption types and parameter ranges in CPCCD capture the relevant corruptions found in real indoor scans
    Used to justify the benchmark's validity (Section 3). No quantitative validation against ScanObjectNN corruption statistics is provided.
  • domain assumption Chamfer Distance, F-score, and Fidelity are adequate metrics for completion quality under corruption
    Metrics adopted from prior work (Section 5.2); CD can be insensitive to outliers.
  • domain assumption Clean partial point cloud features provide a valid supervisory target for denoising in the NMM
    The Positive Loss aligns clean features to clean partial ground truth (Eq. 5); assumes this supervision transfers to inference when ground truth is absent.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Denoising-While-Completing Network (DWCNet): Robust Point Cloud Completion Under Corruption." pith.science (2026). https://pith.science/paper/CWKF5XOC

@misc{pith2026250716743,
  author       = {Pith},
  title        = {Pith review of: Denoising-While-Completing Network (DWCNet): Robust Point Cloud Completion Under Corruption},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWKF5XOC}},
  note         = {Machine review of arXiv:2507.16743}
}
read the original abstract

Point cloud completion is crucial for 3D computer vision tasks in autonomous driving, augmented reality, and robotics. However, obtaining clean and complete point clouds from real-world environments is challenging due to noise and occlusions. Consequently, most existing completion networks -- trained on synthetic data -- struggle with real-world degradations. In this work, we tackle the problem of completing and denoising highly corrupted partial point clouds affected by multiple simultaneous degradations. To benchmark robustness, we introduce the Corrupted Point Cloud Completion Dataset (CPCCD), which highlights the limitations of current methods under diverse corruptions. Building on these insights, we propose DWCNet (Denoising-While-Completing Network), a completion framework enhanced with a Noise Management Module (NMM) that leverages contrastive learning and self-attention to suppress noise and model structural relationships. DWCNet achieves state-of-the-art performance on both clean and corrupted, synthetic and real-world datasets. The dataset and code will be publicly available at https://github.com/keneniwt/DWCNET-Robust-Point-Cloud-Completion-against-Corruptions

Figures

Figures reproduced from arXiv: 2507.16743 by the authors.

Figure 1
Figure 1. We propose a corrupted point cloud completion benchmark [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Real world scans (like ScanObjectNN dataset) often contain [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. CPCCD Dataset and its different types of corruptions Point clouds from active acquisition are more suscepti￾ble to noise and corruption due to factors like the ac￾quisition environment, measurement devices, or human error [24], in contrast to those acquired passively. Datasets in Point Cloud Completion: Due to chal￾lenges in active acquisition, researchers often use pas￾sively acquired, clean, and synthetic pairs of… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Overall architecture of DWCNet. The input point cloud is first down-sampled and local features are extracted using DGCNN [42], [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Noise Managment Module (NMM) attention heads to capture global dependencies while the FFN refines features per-position. Given an input fea￾ture fi , fi ∈ R B×L×D where B is the batch size, L is the sequence length D is the feature dimension, and + represents addition.…
Figure 6
Figure 6. Figure 6: Example of results on CPCCD dataset before fine-tuning: chair [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Example of results on CPCCD dataset before fine-tuning: multiple categories [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Example of results on CPCCD dataset after fine-tuning: chair [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Example of results on CPCCD dataset after fine-tuning: multiple categories [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Some examples of results on ScanObjectNN dataset [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

  1. [1]

    Y . Guo, H. Wang, Q. Hu, H. Liu, L. Liu, M. Bennamoun, Deep learning for 3D point clouds: A survey, PAMI 43 (2020) 4338– 4364

    work page 2020

  2. [2]

    K. W. Tesema, L. Hill, M. W. Jones, M. I. Ahmad, G. K. Tam, Point cloud completion: A survey, IEEE Transactions on Visualization and Computer Graphics 30 (2023) 6880–6899. doi:10.1109/TVCG.2023.3344935. 14

    work page arXiv 2023

  3. [3]

    C. R. Qi, O. Litany, K. He, L. J. Guibas, Deep hough voting for 3D object detection in point clouds, in: ICCV , 2019, pp. 9277–9286

    work page 2019

  4. [4]

    Zheng, Y

    Y . Zheng, Y . Li, S. Yang, H. Lu, Global-pbnet: a novel point cloud registration for autonomous driving, T-ITS 23 (2022) 22312–22319

    work page 2022

  5. [5]

    Varley, C

    J. Varley, C. DeChant, A. Richardson, J. Ruales, P. Allen, Shape completion enabled robotic grasping, in: IROS, 2017, pp. 2442– 2447

    work page 2017

  6. [6]

    Jakli ˇc, M

    A. Jakli ˇc, M. Eri ˇc, I. Mihajlovi ´c, ˇZ. Stopinˇsek, F. Solina, V ol- umetric models from 3D point clouds: The case study of sar- cophagi cargo from a 2nd/3rd century ad roman shipwreck near sutivan on island braˇc, croatia, J. Archaeol. Sci 62 (2015) 143– 152

    work page 2015

  7. [7]

    P. Cai, D. Scott, X. Li, S. Wang, Orthogonal dictionary guided shape completion network for point cloud, in: AAAI, vol- ume 38, 2024, pp. 864–872

    work page 2024

  8. [8]

    X. Yu, Y . Rao, Z. Wang, J. Lu, J. Zhou, AdaPoinTr: Diverse point cloud completion with adaptive geometry-aware trans- formers, arXiv:2301.04545 (2023)

    work page Pith review arXiv 2023

Show all 56 references
  1. [9]

    W. Yuan, T. Khot, D. Held, C. Mertz, M. Hebert, PCN: Point completion network, in: 3DV , 2018, pp. 728–737

    2018

  2. [10]

    L. Pan, X. Chen, Z. Cai, J. Zhang, H. Zhao, S. Yi, Z. Liu, Vari- ational relational point completion network, in: CVPR, 2021, pp. 8524–8533

    2021

  3. [11]

    Geiger, P

    A. Geiger, P. Lenz, C. Stiller, R. Urtasun, Vision meets robotics: The kitti dataset, IJRR (2013)

    2013

  4. [12]

    A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, et al., Shapenet: An information-rich 3D model repository, arXiv:1512.03012 (2015)

    2015 arXiv

  5. [13]

    L. P. Tchapmi, V . Kosaraju, H. Rezatofighi, I. Reid, S. Savarese, Topnet: Structural point cloud decoder, in: CVPR), 2019, pp. 383–392

    2019

  6. [14]

    Y . Zhou, F. Tu, K. Sha, J. Ding, H. Chen, A survey on data qual- ity dimensions and tools for machine learning, arXiv preprint arXiv:2406.19614 (2024)

    2024 arXiv

  7. [15]

    J. Ren, L. Pan, Z. Liu, Benchmarking and analyzing point cloud classification under corruptions, in: Proceedings of the 39th International Conference on Machine Learning, volume 162, 2022, pp. 18559–18575

    2022

  8. [16]

    Arora, S

    H. Arora, S. Mishra, S. Peng, K. Li, A. Mahdavi-Amiri, Multi- modal shape completion via imle, arXiv:2106.16237 (2021)

    2021 arXiv

  9. [17]

    L. Kong, Y . Liu, X. Li, R. Chen, W. Zhang, J. Ren, L. Pan, K. Chen, Z. Liu, Robo3D: Towards robust and reliable 3D per- ception against corruptions, in: Proceedings of the IEEE /CVF International Conference on Computer Vision, 2023, pp. 19994– 20006

    2023

  10. [18]

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, M. Nießner, Scannet: Richly-annotated 3D reconstructions of indoor scenes, in: CVPR, 2017, pp. 5828–5839

    2017

  11. [19]

    R. Cui, S. Qiu, S. Anwar, J. Liu, C. Xing, J. Zhang, N. Barnes, P2c: Self-supervised point cloud completion from single partial clouds, arXiv:2307.14726 (2023)

    2023 arXiv

  12. [20]

    J. Wang, Y . Cui, D. Guo, J. Li, Q. Liu, C. Shen, Pointattn: You only need attention for point cloud completion, arXiv:2203.08485 (2022)

    2022 arXiv

  13. [21]

    M. A. Uy, Q.-H. Pham, B.-S. Hua, D. T. Nguyen, S.-K. Yeung, Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data, in: International Conference on Computer Vision (ICCV), 2019

    2019

  14. [22]

    Q. Wang, Y . Tan, Z. Mei, Computational methods of acquisition and processing of 3D point cloud data for construction applica- tions, Arch. Computat. Methods Eng 27 (2019) 479–499

    2019

  15. [23]

    L. Li, R. Wang, X. Zhang, A tutorial review on point cloud reg- istrations: Principle, classification, comparison, and technology challenges, Math. Probl. Eng 2021 (2021)

    2021

  16. [24]

    J. Ren, L. Pan, Z. Liu, Benchmarking and analyzing point cloud classification under corruptions, International Conference on Machine Learning (ICML) (2022)

    2022

  17. [25]

    Y . Dong, C. Kang, J. Zhang, Z. Zhu, Y . Wang, X. Yang, H. Su, X. Wei, J. Zhu, Benchmarking robustness of 3D object detec- tion to common corruptions, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 1022–1032

    2023

  18. [26]

    Gurumurthy, S

    S. Gurumurthy, S. Agrawal, High fidelity semantic shape com- pletion for point clouds using latent optimization, in: W ACV , 2019, pp. 1099–1108

    2019

  19. [27]

    J. Li, G. Si, X. Liang, Z. An, P. Tian, F. Zhou, Partition-based point cloud completion network with density refinement, En- tropy 25 (2023) 1018

    2023

  20. [28]

    C. Ma, Y . Yang, J. Guo, C. Wang, Y . Guo, Completing par- tial point clouds with outliers by collaborative completion and segmentation, arXiv:2203.09772 (2022)

    2022 arXiv

  21. [29]

    R.-W. Li, B. Wang, C.-P. Li, L.-X. Zhang, L. Gao, High-fidelity point cloud completion with low-resolution recovery and noise- aware upsampling, arXiv:2112.11271 (2021)

    2021 arXiv

  22. [31]

    L. Zhou, G. Sun, Y . Li, W. Li, Z. Su, Point cloud denoising re- view: from classical to deep learning-based approaches, Graph- ical Models 121 (2022) 101140

    2022

  23. [32]

    S. Luo, W. Hu, Score-based point cloud denoising, in: Proceed- ings of the IEEE /CVF International Conference on Computer Vision, 2021, pp. 4583–4592

    2021

  24. [33]

    Z. Wei, H. Chen, L. Nan, J. Wang, J. Qin, M. Wei, Pathnet: Path- selective point cloud denoising, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (2024) 4426–4442

    2024

  25. [34]

    G. Hou, G. Qin, M. Sun, Y . Liang, J. Yan, Z. Zhang, Noise- trans: point cloud denoising with transformers, arXiv preprint arXiv:2304.11812 (2023)

    2023 arXiv

  26. [35]

    Hermosilla, T

    P. Hermosilla, T. Ritschel, T. Ropinski, Total denoising: Unsu- pervised learning of 3d point cloud cleaning, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 52–60

    2019

  27. [36]

    S. Xie, J. Gu, D. Guo, C. R. Qi, L. Guibas, O. Litany, Point- contrast: Unsupervised pre-training for 3d point cloud under- standing, in: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, Springer, 2020, pp. 574–591

    2020

  28. [37]

    Z. Pan, G. Liu, W. Gao, T. H. Li, Epcontrast: e ffective point- level contrastive learning for large-scale point cloud understand- ing, in: 2024 IEEE International Conference on Multimedia and Expo (ICME), IEEE, 2024, pp. 1–6

    2024

  29. [38]

    A. Du, T. Zhou, S. Pang, Q. Wu, J. Zhang, Pcl: Point con- trast and labeling for weakly supervised point cloud semantic segmentation, IEEE Transactions on Multimedia (2024)

    2024

  30. [39]

    Z. Wang, Q. Xie, M. Wei, K. Long, J. Wang, Multi-feature fusion votenet for 3d object detection, ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 18 (2022) 1–17

    2022

  31. [40]

    Avetisyan, M

    A. Avetisyan, M. Dahnert, A. Dai, M. Savva, A. X. Chang, M. Nießner, Scan2cad: Learning cad model alignment in rgb-d scans, in: CVPR, 2019, pp. 2614–2623

    2019

  32. [41]

    B. M. ¨Ocal, M. Tatarchenko, S. Karaoglu, T. Gevers, Realdi ff: Real-world 3d shape completion using self-supervised diffusion models, arXiv preprint arXiv:2409.10180 (2024)

    2024 arXiv

  33. [42]

    A. V . Phan, M. Le Nguyen, Y . L. H. Nguyen, L. T. Bui, DGCNN: A convolutional neural network over large-scale la- 15 beled graphs, Neural Networks 108 (2018) 533–543

    2018

  34. [43]

    Y . Hu, C. Chen, H. Zou, X. Zhong, E. S. Chng, Unifying speech enhancement and separation with gradient modulation for end- to-end noise-robust speech separation, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2023...

    2023

  35. [44]

    Sivaraman, M

    A. Sivaraman, M. Kim, E fficient personalized speech enhance- ment through self-supervised learning, IEEE Journal of Selected Topics in Signal Processing 16 (2022) 1342–1356

    2022

  36. [45]

    Kharitonov, M

    E. Kharitonov, M. Rivi `ere, G. Synnaeve, L. Wolf, P.-E. Mazar´e, M. Douze, E. Dupoux, Data augmenting contrastive learning of speech representations in the time domain, in: 2021 IEEE Spo- ken Language Technology Workshop (SLT), IEEE, 2021, pp. 215–222

    2021

  37. [46]

    T. Chen, C. Luo, L. Li, Intriguing properties of contrastive losses, Advances in Neural Information Processing Systems 34 (2021) 11834–11845

    2021

  38. [47]

    Cordonnier, A

    J.-B. Cordonnier, A. Loukas, M. Jaggi, Multi-head atten- tion: Collaborate instead of concatenate, arXiv preprint arXiv:2006.16362 (2020)

    2020 arXiv

  39. [48]

    Zhang, M

    O. Zhang, M. Wu, J. Bayrooti, N. Goodman, Tempera- ture as uncertainty in contrastive learning, arXiv preprint arXiv:2110.04403 (2021)

    2021 arXiv

  40. [49]

    Y . Yang, C. Feng, Y . Shen, D. Tian, Foldingnet: Point cloud auto-encoder via deep grid deformation, in: CVPR, 2018, pp. 206–215

    2018

  41. [50]

    H. Xie, H. Yao, S. Zhou, J. Mao, S. Zhang, W. Sun, Grnet: Gridding residual network for dense point cloud completion, in: EECV , 2020, pp. 365–381

    2020

  42. [51]

    X. Yu, Y . Rao, Z. Wang, Z. Liu, J. Lu, J. Zhou, Pointr: Diverse point cloud completion with geometry-aware transformers, in: ICCV , 2021, pp. 12498–12507

    2021

  43. [52]

    Xiang, X

    P. Xiang, X. Wen, Y .-S. Liu, Y .-P. Cao, P. Wan, W. Zheng, Z. Han, Snowflakenet: Point cloud completion by snowflake point deconvolution with skip-transformer, in: ICCV , 2021, pp. 5499–5509

    2021

  44. [53]

    Nguyen, Q.-H

    T. Nguyen, Q.-H. Pham, T. Le, T. Pham, N. Ho, B.-S. Hua, Point-set distances for learning representations of 3D point clouds, in: ICCV , 2021, pp. 10478–10487

    2021

  45. [54]

    T. Wu, L. Pan, J. Zhang, T. Wang, Z. Liu, D. Lin, Density- aware chamfer distance as a comprehensive metric for point cloud completion, arXiv:2111.12702 (2021)

    2021 arXiv

  46. [55]

    Moravec, A comparative study: L1-norm vs

    J. Moravec, A comparative study: L1-norm vs. l2-norm; point- to-point vs. point-to-line metric; evolutionary computation vs. gradient search, Applied Artificial Intelligence 29 (2015) 164– 210

    2015

  47. [56]

    Sokolova, N

    M. Sokolova, N. Japkowicz, S. Szpakowicz, Beyond accuracy, f-score and roc: a family of discriminant measures for perfor- mance evaluation, in: Adv. in Artif. Intell., 2006, pp. 1015– 1021

    2006

  48. [57]

    L. Jing, P. Vincent, Y . LeCun, Y . Tian, Understanding dimen- sional collapse in contrastive self-supervised learning, arXiv preprint arXiv:2110.09348 (2021). Appendix A. Quantitative Results for the Robust- ness Test Appendix A.1. Before Fine-tuning In this section we presen...

    2021 arXiv

Pith tools

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