Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Hybrid Long and Short Range Flows for Point Cloud Filtering

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read HybridPF claims that combining short-range score and long-range flow information improves point cloud denoising, reducing clustering and overshoot.

desk verdict Solid hybrid-flow point cloud filter whose high-noise gains depend on test-time tuned step size; deserves review but needs code and a less overstated generalization claim. read the letter →

arxiv 2508.08542 v1 pith:62XYEG6F submitted 2025-08-12 cs.GR cs.CV

Dasith de Silva Edirimuni , Xuequan Lu , Ajmal Saeed Mian , Lei Wei , Gang Li , Scott Schaefer , Ying He This is my paper
classification cs.GRcs.CV
keywords pointcloudfilteringdenoisingscorematchingrectifiedflowdynamicgraphconvolutionChamferdistancelong-rangeshort-range
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

The paper tries to establish that point cloud denoising is improved by combining two flow estimates rather than relying on either alone: short-range score displacements, which push noisy points toward the clean surface, and long-range velocity flow, which carries the overall displacement from a heavily noised patch to its clean counterpart. The proposed HybridPF network trains a LongModule and a ShortModule in parallel, feeds long-range encoder features into the short-range score branch, and decodes with dynamic graph convolutional layers instead of plain fully connected layers. If this is right, filtered point clouds should land closer to the true surface and keep a more even point distribution than score-only or flow-only methods. The paper reports exactly that: lower Chamfer and point-to-mesh errors than prior methods on synthetic and real scans, with fewer clustering artifacts and faster inference.

What carries the argument

Two mechanisms carry the argument. First, hybrid conditioning: the LongModule encoder's latent features, trained to estimate the long-range velocity $\mathbf{v}(\mathbf{x}_t) = \mathbf{x}_1-\mathbf{x}_0$, are fed into the ShortModule's score estimator, so the short-range update $\mathbf{x}_1-\mathbf{x}_t$ is informed by where the trajectory started and where it should end. Second, a Dynamic EdgeConv decoder that builds a k-nearest-neighbor graph among point features in latent space during decoding, replacing the fully connected decoders used by prior displacement methods. The graph decoder is what lets the network use local feature neighborhoods when converting high-dimensional features to 3

What would settle it

Filter the PUNet sigma=3% test clouds with the same trained model but keep the low-noise step alpha=0.8, without manual rescaling. If Chamfer distance stays at the reported level, the high-noise success is not due to test-time alpha tuning; if CD jumps toward or above StraightPCF's level, the generalization claim depends on knowing sigma. A second check: replace the Dynamic EdgeConv decoder with an FC decoder at equal parameter count and measure CD; a vanishing gap would falsify the decoder's contribution.

Watch

Extended reading notes

Core claim

The central claim is that a filter does better when its short-range and long-range trajectory estimates are trained jointly and the long-range ones are allowed to steer the short-range ones. HybridPF trains a LongModule to predict the constant displacement $\mathbf{x}_1-\mathbf{x}_0$ from an intermediate noisy patch $\mathbf{x}_t$, and a ShortModule to predict the score displacement $\mathbf{x}_1-\mathbf{x}_t$ conditional on the LongModule encoder's features. At inference the LongModule decoder is discarded, and the ShortModule is iterated four times with step size $\alpha$. The paper argues this hybrid conditioning, plus a dynamic graph convolutional decoder, removes the clustering that sco

Load-bearing premise

The load-bearing premise is that the input noise level is known at test time, because the reported high-noise results are obtained by manually increasing the step size alpha (e.g., alpha=1.3 at sigma=3%, 10K points); if the noise level is unknown or misestimated, the displacement scaling in Eq. (11) is miscalibrated and the claimed high-noise generalization is not established.

Editorial extensions

If this is right

  • On PUNet and PCNet with Gaussian noise, HybridPF matches or exceeds the best competing method on Chamfer distance at nearly every resolution and noise level, including 3% noise unseen in training.
  • Combining long-range conditioning with the score branch produces better point distributions and fewer clustering artifacts than score-only baselines.
  • The dynamic graph convolutional decoder contributes measurable gains over fully connected decoders at equal parameter count, as shown by the HybridPFv1 ablation.
  • Inference is faster than IterativePFN and StraightPCF because high noise does not require repeating the entire filtering run; only the step size changes.
  • The method generalizes beyond Gaussian noise, achieving best or second-best results on non-isotropic, uniform, and Laplace noise patterns in the supplementary experiments.

Reading between the lines

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

  • The reported dependence on a hand-set step size for high noise suggests the method's generalization is partly calibration, not purely learned; a natural extension is to predict the displacement scale from the input patch so the noise level need not be known at test time.
  • The hybrid flow idea is generic: any trajectory task with a coarse-to-fine map, such as point cloud upsampling or generation, could condition short-range updates on long-range displacement features.
  • The paper's logic implies hybrid conditioning should matter most where long-range information is least ambiguous, namely high noise and sparse clouds; the growing ablation gap from 1% to 3% noise is consistent with that hypothesis but is not tested as a separate claim.
  • Because the LongModule decoder is discarded at inference, it acts as a training-time regularizer; one could test whether distilling the long-range features alone or removing the decoder entirely changes the final filtering quality.
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

3 major / 5 minor

Summary. The paper proposes HybridPF, a point cloud filtering method that combines a ShortModule (score/displacement estimation) and a LongModule (straight-flow displacement estimation), where long-range features from the LongEncoder condition the ShortModule. The authors also replace fully connected decoders with dynamic graph convolutional layers. Training uses a joint EMD-plus-flow loss; at inference, iterative updates with step size alpha are applied. Experiments are reported on PUNet, PCNet, Kinect, and Paris-Rue-Madame data under Gaussian, non-isotropic Gaussian, uniform, and Laplace noise, with the method claimed to achieve state-of-the-art Chamfer distances and competitive Point2Mesh distances at comparable or better runtime. The supplementary material describes ablations and test-time tuning of alpha per noise level and resolution.

Significance. If the reported gains hold under a fixed evaluation protocol, the hybrid conditioning and dynamic graph decoder are useful contributions to point cloud filtering. The paper's ablations are matched in parameter count and show consistent improvements from the LongModule and the Dynamic EdgeConv decoder, and the evaluation breadth across multiple datasets and noise patterns is a strength. The main weakness is that the high-noise state-of-the-art claim depends on test-time alpha selection rather than a fixed or self-calibrating schedule, and no error bars or code are provided. The long-range 'guidance' is also not an independent information source: both modules are trained on the same (x0, x1) pairs and the LongEncoder feature is a deterministic function of x_t. The empirical gains are still meaningful, but the interpretation and the generalization claim need to be qualified.

major comments (3)
  1. [Sec. IV-B, Eq. (10)-(11), Supplementary Sec. A] The sigma=3% results are obtained with alpha=1.3 at 10K and alpha=1.5 at 50K, while sigma<=2% uses alpha=0.8. Since the ShortModule has no noise-level input and is trained only with sigma_H=2%, the network cannot self-calibrate to 3% noise; the reported high-noise gains are partly an artifact of test-time calibration to the known noise level. The claim that HybridPF 'generalizes well to high noise levels unseen during training' is therefore not established for unknown or misestimated sigma. Please report results with a single alpha, or provide a principled/self-calibrating scaling rule, and include a sensitivity analysis of alpha. If this is not feasible, the generalization claim should be weakened accordingly.
  2. [Table I and Table IV] Headline differences are small, e.g., PUNet 50K sigma=3% CD 1.27 vs 1.31 for StraightPCF, and PUNet 10K sigma=1% CD is tied with PD-LTS at 1.83. No error bars, per-shape breakdowns, or significance tests are reported. Moreover, the P2M advantage is not uniform: at PUNet 50K sigma=3%, DeepPSR achieves 0.58 vs 0.61 for Ours; at 10K sigma=3%, PD-LTS achieves 1.10 vs 1.11 for Ours. The claim of 'best CD results' is supported, but 'almost all metrics' is overstated. Please provide variance or statistical evidence and qualify the claim accordingly.
  3. [Sec. III-D and III-E] The conceptual claim that 'long-range information guides short-range scores' is not directly demonstrated. The LongModule and ShortModule are trained on the same (x0, x1) pairs, and the conditioning feature E_L(x_t; phi1) is a deterministic function of the same input x_t; thus the conditioning is a learned feature, not an independent source of information about x1. The ablation does support the empirical benefit of the architecture, but the paper should avoid implying that long-range information is external to the score estimator. In addition, Eq. (10)-(11) set the input noisy patch as x_tilde_0 even though 'the starting time for the filtering process is not known'; this needs clarification because the network is trained on x_t at variable t.
minor comments (5)
  1. [Abstract / Table III] The abstract claims 'faster inference speed,' but Table III shows PD-LTS at 13.46s vs Ours at 15.56s. Please specify that the improvement is over IterativePFN and StraightPCF rather than all baselines.
  2. [Eq. (3)] The target (x1 - x_t) is called a score, but the score would be (x1 - x_t)/sigma^2 under Eq. (2). Please clarify the notation or explicitly state that the network predicts a displacement rather than a normalized score.
  3. [Sec. VI] Typo: 'close to the the original clean points' should read 'close to the original clean points'.
  4. [Supplementary Sec. A] The supplement labels both the overall section and the first subsection as 'A'. Please renumber to avoid confusion.
  5. [General] No code or checkpoints are provided. Given the small metric differences that constitute the headline results, releasing code would materially aid verification.

Circularity Check

1 steps flagged · score 4.0 of 10

High-noise state-of-the-art results are partly produced by a test-time fitted step size α, so the 'generalization to unseen high noise' claim reduces in part to calibration; the hybrid-flow core is otherwise independently evaluated.

  1. fitted input called prediction [Supplementary Sec. A ('Discussion on the discretization parameter α and iteration number N'); Eq. (10)-(11); Table I]
    "the network overspecializes on the training noise scales, where the maximum noise level is σ = 2% . Therefore, for filtered displacements of point clouds at higher noise, we can simply vary α at test time to scale the inferred displacements. In general, for 0% ≤ σ ≤ 2%, we set α = 0.8 based on empirical results. This procedure is similar to the way in which ScoreDenoise [15] and DeepPSR [17] obtain α. For σ = 3% and 10K resolution, we set α = 1.3. Similarly, at σ = 3% and 50K resolution, we set α = 1.5."

    Eq. (11) sets the filtered position as x~1 = x~0 + α Σ_t s(...); α multiplies the whole learned displacement. The supplementary states the network 'overspecializes' on σ≤2% and that one can 'simply vary α at test time' to 1.3 (10K) or 1.5 (50K) for σ=3%. Since the model is trained only at σ_H=2% and has no noise-level input, the σ=3% results in Table I are produced by scaling the output to the known test noise, i.e., a fitted test-time constant is presented as generalization to unseen high noise. The remainder of the method is held-out evaluated against external baselines, so the circularity is partial.

full rationale

The central derivation is not otherwise circular. Eqs. (3), (6)-(9) restate the standard score-matching and rectified-flow objectives; the long-range flow (x1-x0) and short-range score (x1-x_t) are both defined from the same clean/noisy pairs and are algebraically linked by Eq. (6) (x1-x_t = (1-t)(x1-x0)), but this is an acknowledged modeling choice rather than a hidden equivalence of the claimed result to its inputs. The citations of the authors' own IterativePFN [16] and StraightPCF [22] are not load-bearing in a circular sense: the needed equations appear in the paper itself and the comparisons are to externally reported numbers. The one concrete reduction is the test-time α calibration: the high-noise claim is partly forced by choosing α per noise level/resolution. This raises the score to 4 but not higher, because the main empirical contribution (hybrid guidance plus graph decoder, tested at σ=1%,2% and on unseen datasets) retains independent content and would stand even if the σ=3% calibration were removed.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper builds on standard score-matching and rectified-flow assumptions, fits several hyperparameters to validation or test conditions (notably alpha per noise level), and assumes cross-noise generalization from a single training distribution. No invented physical or mathematical entities are introduced; the 'long-range feature' is a learned latent representation, not a postulated entity.

free parameters (5)
  • alpha (inference step size) = 0.8 for sigma<=2%, 1.3 for sigma=3% 10K, 1.5 for sigma=3% 50K
    Hand-set at test time per noise level to scale filtered displacements in Eq. (10)-(11). Reported in Supplementary Sec. A.
  • lambda (loss weight) = 10.0
    Weights EMD short-range loss vs long-range L2 loss in Eq. (9). Chosen via ablation in Supplementary Table VI.
  • N (number of filtering iterations) = 4
    Number of score application steps in Eq. (10)-(11). Chosen via ablation in Supplementary Table V.
  • sigma_H (training noise scale) = 2% of bounding sphere radius
    Standard deviation used to construct high-noise variants x0 in Eq. (6) and during training.
  • k (graph neighbors) = 32
    Nearest neighbors used in the encoder's dynamic graph convolution (Sec III-A).
assumptions (5)
  • domain assumption Reverse SDE score matching (Eq. 2) provides the filtering displacement grad_x log p(x_t) = (x_1 - x_t)/sigma^2.
    Invoked in Sec III-B as 'well established' from Song et al. [18,19]; underpins the ShortModule objective without re-derivation.
  • domain assumption Intermediate noisy patches satisfy linear interpolation x_t = (1-t)x_0 + t x_1 (Eq. 6).
    Adopted from rectified flow literature [20,21] in Sec III-C; the LongModule's target x_1 - x_0 is only constant under this interpolation.
  • domain assumption x_0 and x_1 are independent and identically distributed samples from Pi_0 and Pi_1.
    Stated in Sec III-C; required for the long-range flow loss Eq. (8) to be well-posed.
  • domain assumption Training on PUNet with isotropic Gaussian noise at sigma_H=2% transfers to unseen noise types (non-isotropic, uniform, Laplace, Kinect, LiDAR).
    The paper only trains on Gaussian PUNet data, yet claims SOTA on other noise patterns and real-world scans; this generalization is empirically assumed rather than derived.
  • domain assumption Dynamic graph convolution (DGCNN [13]) is an effective encoder and decoder backbone for point features.
    The encoder and proposed decoder rely on DGCNN message passing from Sec III-A; treated as background without theoretical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Long and Short Range Flows for Point Cloud Filtering." pith.science (2026). https://pith.science/paper/62XYEG6F

@misc{pith2026250808542,
  author       = {Pith},
  title        = {Pith review of: Hybrid Long and Short Range Flows for Point Cloud Filtering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/62XYEG6F}},
  note         = {Machine review of arXiv:2508.08542}
}
abstract

Point cloud capture processes are error-prone and introduce noisy artifacts that necessitate filtering/denoising. Recent filtering methods often suffer from point clustering or noise retaining issues. In this paper, we propose Hybrid Point Cloud Filtering ($\textbf{HybridPF}$) that considers both short-range and long-range filtering trajectories when removing noise. It is well established that short range scores, given by $\nabla_{x}\log p(x_t)$, may provide the necessary displacements to move noisy points to the underlying clean surface. By contrast, long range velocity flows approximate constant displacements directed from a high noise variant patch $x_0$ towards the corresponding clean surface $x_1$. Here, noisy patches $x_t$ are viewed as intermediate states between the high noise variant and the clean patches. Our intuition is that long range information from velocity flow models can guide the short range scores to align more closely with the clean points. In turn, score models generally provide a quicker convergence to the clean surface. Specifically, we devise two parallel modules, the ShortModule and LongModule, each consisting of an Encoder-Decoder pair to respectively account for short-range scores and long-range flows. We find that short-range scores, guided by long-range features, yield filtered point clouds with good point distributions and convergence near the clean surface. We design a joint loss function to simultaneously train the ShortModule and LongModule, in an end-to-end manner. Finally, we identify a key weakness in current displacement based methods, limitations on the decoder architecture, and propose a dynamic graph convolutional decoder to improve the inference process. Comprehensive experiments demonstrate that our HybridPF achieves state-of-the-art results while enabling faster inference speed.

Figures

Figures reproduced from arXiv: 2508.08542 by the authors.

Figure 1
Figure 1. We visualize the hybrid flows which recover better distributions of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We use t-SNE to visualize features of a single point, at different noise levels. At lower noise levels, we expect features to be similar (i.e., small distance between darker points). As noise increases, features are less consistent. However, the encoder should ideally distinguish features across noise levels (clearer separation of colors). The baseline score model is sub-optimal as features at lower noise (dark) are… view at source ↗
Figure 3
Figure 3. Our HybridPF architecture. It comprises of a LongModule that infers long range flow and a ShortModule that infers short range flow. During training, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Heatmap visualization of filtered point clouds on PUNet and PCNet data at high resolution and high Gaussian noise (i.e., 50K resolution shapes with [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of filtering results on two scenes of the Paris-Rue-Madame dataset. We observe more significant clustering artifacts produced by other [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Heatmap visualization of filtered point clouds on PUNet and PCNet data at high resolution and high non-isotropic Gaussian noise (i.e., 50K resolution [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Heatmap visualization of filtered point clouds on PUNet and PCNet data at high resolution and high uniformly-distributed noise (i.e., 50K resolution [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Heatmap visualization of filtered point clouds on PUNet and PCNet data at high resolution and high Laplace noise (i.e., 50K resolution shapes with [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 48 canonical work pages

  1. [1]

    Mesh-offset-based method to generate a delta volume to support the maintenance of partially damaged parts through 3d printing,

    Y . Kim, K. Kwon, and D. Mun, “Mesh-offset-based method to generate a delta volume to support the maintenance of partially damaged parts through 3d printing,” Journal of Mechanical Science and Technology , vol. 35, no. 7, pp. 3131–3143, 2021

    work page 2021

  2. [2]

    Point cloud modeling as a bridge between landscape design and planning,

    P. R. W. Urech, M. Dissegna, C. Girot, and A. Grêt-Regamey, “Point cloud modeling as a bridge between landscape design and planning,” Landscape and Urban Planning , vol. 203, p. 103903, 2020

    work page 2020

  3. [3]

    Self-supervised pillar motion learning for autonomous driving,

    C. Luo, X. Yang, and A. Yuille, “Self-supervised pillar motion learning for autonomous driving,” in CVPR, 2021, Conference Proceedings

    work page 2021

  4. [4]

    Feature preserving point set surfaces based on non linear kernel regression,

    A. C. Öztireli, G. Guennebaud, and M. Gross, “Feature preserving point set surfaces based on non linear kernel regression,” Computer Graphics Forum, vol. 28, 2009

    work page 2009

  5. [5]

    Bilateral mesh denoising,

    S. Fleishman, I. Drori, and D. Cohen-Or, “Bilateral mesh denoising,” ACM SIGGRAPH 2003 Papers , 2003

    work page 2003

  6. [6]

    Semreg: Semantics constrained point cloud registration,

    S. Fung, X. Lu, D. d. S. Edirimuni, W. Pan, X. Liu, and H. Li, “Semreg: Semantics constrained point cloud registration,” in Computer Vision – ECCV 2024. Cham: Springer Nature Switzerland, 2025, pp. 293–310

    work page 2024

  7. [7]

    KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d,

    Y . Liao, J. Xie, and A. Geiger, “KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d,” arXiv preprint arXiv:2109.13410, 2021

    arXiv 2021

  8. [8]

    The bilateral filter for point clouds,

    J. Digne and C. D. Franchis, “The bilateral filter for point clouds,” Image Process. Line, vol. 7, pp. 278–287, 2017

    work page 2017

Show all 51 references
  1. [9]

    Point cloud denoising via moving rpca,

    E. Mattei and A. Castrodad, “Point cloud denoising via moving rpca,” Computer Graphics Forum, vol. 36, pp. 123–137, 2017

    2017

  2. [10]

    Feature graph learning for 3d point cloud denoising,

    W. Hu, X. Gao, G. Cheung, and Z. Guo, “Feature graph learning for 3d point cloud denoising,” IEEE Transactions on Signal Processing , vol. 68, pp. 2841–2856, 2020

    2020

  3. [11]

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

    C. Qi, H. Su, K. Mo, and L. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 77–85, 2017

    2017

  4. [12]

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

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” in Advances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017, Conference Proceedings

    2017

  5. [13]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” ACM Transactions on Graphics (TOG) , 2019

    2019

  6. [14]

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

    M.-J. Rakotosaona, V . L. Barbera, P. Guerrero, N. Mitra, and M. Ovs- janikov, “Pointcleannet: Learning to denoise and remove outliers from dense point clouds,” Computer Graphics Forum, vol. 39, 2020

    2020

  7. [15]

    Score-based point cloud denoising,

    S. Luo and W. Hu, “Score-based point cloud denoising,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , October 2021, pp. 4583–4592

    2021

  8. [16]

    Iterativepfn: True iterative point cloud filtering,

    D. de Silva Edirimuni, X. Lu, Z. Shao, G. Li, A. Robles-Kelly, and Y . He, “Iterativepfn: True iterative point cloud filtering,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 6 2023, pp. 13 530–13 539

    2023

  9. [17]

    Deep point set resampling via gradient fields,

    H. Chen, B. Du, S. Luo, and W. Hu, “Deep point set resampling via gradient fields,” IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 45, pp. 2913–2930, 3 2023

    2023

  10. [18]

    Generative modeling by estimating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by estimating gradients of the data distribution,” in Proceedings of the 33rd International Conference on Neural Information Processing Systems . Curran Associates Inc., 2019, Conference Proceedings

    2019

  11. [19]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” in International Conference on Learning Representations, 2021. [Online]. Available: https://openreview.net/forum? id=PxTIG12RRHS

    2021

  12. [20]

    Flow straight and fast: Learning to generate and transfer data with rectified flow,

    X. Liu, C. Gong, and Q. Liu, “Flow straight and fast: Learning to generate and transfer data with rectified flow,” in The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . OpenReview.net, 2023

    2023

  13. [21]

    Fast point cloud generation with straight flows,

    L. Wu, D. Wang, C. Gong, X. Liu, Y . Xiong, R. Ranjan, R. Krishnamoor- thi, V . Chandra, and Q. Liu, “Fast point cloud generation with straight flows,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 6 2023, pp. 9445–9454

    2023

  14. [22]

    Straightpcf: Straight point cloud filtering,

    D. de Silva Edirimuni, X. Lu, G. Li, L. Wei, A. Robles-Kelly, and H. Li, “Straightpcf: Straight point cloud filtering,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024 . IEEE, 2024, pp. 20 721–20 730

    2024

  15. [23]

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

    A. Mao, Z. Du, Y .-H. Wen, J. Xuan, and Y .-J. Liu, “Pd-flow: A point cloud denoising framework with normalizing flows,” in The European Conference on Computer Vision (ECCV) , 2022, Conference Proceedings

    2022

  16. [24]

    The approximation power of moving least-squares,

    D. Levin, “The approximation power of moving least-squares,” Math. Comput., vol. 67, pp. 1517–1531, 1998

    1998

  17. [25]

    Computing and rendering point set surfaces,

    M. Alexa, J. Behr, D. Cohen-Or, S. Fleishman, D. Levin, and C. T. Silva, “Computing and rendering point set surfaces,” IEEE Trans. Vis. Comput. Graph., vol. 9, pp. 3–15, 2003

    2003

  18. [26]

    Point-sampled cell complexes,

    A. Adamson and M. Alexa, “Point-sampled cell complexes,” ACM Trans. Graph., vol. 25, pp. 671–680, 2006

    2006

  19. [27]

    Algebraic point set surfaces,

    G. Guennebaud and M. Gross, “Algebraic point set surfaces,” in SIGGRAPH 2007, 2007, Conference Proceedings

    2007

  20. [28]

    Surface reconstruction from unorganized points,

    H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, and W. Stuetzle, “Surface reconstruction from unorganized points,” Proceedings of the 19th annual conference on Computer graphics and interactive techniques , 1992

    1992

  21. [29]

    Estimating differential quantities using polynomial fitting of osculating jets,

    F. Cazals and M. Pouget, “Estimating differential quantities using polynomial fitting of osculating jets,” Computer Aided Geometric Design , vol. 22, no. 2, pp. 121–146, 2005

    2005

  22. [30]

    Similarity based filtering of point clouds,

    J. Digne, “Similarity based filtering of point clouds,” 2012 IEEE Com- puter Society Conference on Computer Vision and Pattern Recognition Workshops, pp. 73–79, 2012

    2012

  23. [31]

    L1-sparse reconstruction of sharp point set surfaces,

    H. Avron, A. Sharf, C. Greif, and D. Cohen-Or, “L1-sparse reconstruction of sharp point set surfaces,” ACM Trans. Graph. , vol. 29, pp. 135:1– 135:12, 2010

    2010

  24. [32]

    Denoising point sets via l0 minimization,

    Y . Sun, S. Schaefer, and W. Wang, “Denoising point sets via l0 minimization,” Comput. Aided Geom. Des. , vol. 35-36, pp. 2–15, 2015

    2015

  25. [33]

    Low rank matrix approx- imation for 3d geometry filtering,

    X. Lu, S. Schaefer, J. Luo, L. Ma, and Y . He, “Low rank matrix approx- imation for 3d geometry filtering,” IEEE transactions on visualization and computer graphics , vol. PP, 2020

    2020

  26. [34]

    Parameterization- free projection for geometry reconstruction,

    Y . Lipman, D. Cohen-Or, D. Levin, and H. Tal-Ezer, “Parameterization- free projection for geometry reconstruction,” ACM SIGGRAPH 2007 papers, 2007

    2007

  27. [35]

    Consolidation of unorganized point clouds for surface reconstruction,

    H. Huang, D. Li, H. Zhang, U. Ascher, and D. Cohen-Or, “Consolidation of unorganized point clouds for surface reconstruction,” ACM SIGGRAPH Asia 2009 papers , 2009

    2009

  28. [36]

    Continuous projection for fast l1 reconstruction,

    R. Preiner, O. Mattausch, M. Arikan, R. Pajarola, and M. Wimmer, “Continuous projection for fast l1 reconstruction,” ACM Transactions on Graphics (TOG), vol. 33, pp. 1 – 13, 2014

    2014

  29. [37]

    Deep feature-preserving normal estimation for point cloud filtering,

    D. Lu, X. Lu, Y . Sun, and J. Wang, “Deep feature-preserving normal estimation for point cloud filtering,” Computer-Aided Design, vol. 125, 2020

    2020

  30. [38]

    Pointpronets: Consolidation of point clouds with convolutional neural networks,

    R. Roveri, A. C. Öztireli, I. Pandele, and M. Gross, “Pointpronets: Consolidation of point clouds with convolutional neural networks,” Computer Graphics Forum, vol. 37, no. 2, pp. 87–99, 2018

    2018

  31. [39]

    Pointfilter: Point cloud filtering via encoder-decoder modeling,

    D. Zhang, X. Lu, H. Qin, and Y . He, “Pointfilter: Point cloud filtering via encoder-decoder modeling,” IEEE Transactions on Visualization and Computer Graphics, vol. 27, pp. 2015–2027, 2021

    2015

  32. [40]

    Contrastive learning for joint normal estimation and point cloud filtering,

    D. de Silva Edirimuni, X. Lu, G. Li, and A. Robles-Kelly, “Contrastive learning for joint normal estimation and point cloud filtering,” IEEE Trans. Vis. Comput. Graph. , vol. 30, no. 8, pp. 4527–4541, 2024

    2024

  33. [41]

    Learning graph- convolutional representations for point cloud denoising,

    F. Pistilli, G. Fracastoro, D. Valsesia, and E. Magli, “Learning graph- convolutional representations for point cloud denoising,” in Computer Vision – ECCV 2020 . Springer International Publishing, 2020, Conference Proceedings, pp. 103–118. JOURNAL OF LATEX CLASS FILES, VOL. 1...

    2020

  34. [42]

    Differentiable manifold reconstruction for point cloud denoising,

    S. Luo and W. Hu, “Differentiable manifold reconstruction for point cloud denoising,” in Proceedings of the 28th ACM International Conference on Multimedia . Association for Computing Machinery, 2020, Conference Proceedings, p. 1330–1338. [Online]. Available: https://doi.org/1...

    2020

  35. [43]

    Learning gradient fields for shape generation,

    R. Cai, G. Yang, H. Averbuch-Elor, Z. Hao, S. Belongie, N. Snavely, and B. Hariharan, “Learning gradient fields for shape generation,” in Computer Vision – ECCV 2020. Springer International Publishing, 2020, Conference Proceedings, pp. 364–381

    2020

  36. [44]

    Diffusion probabilistic models for 3d point cloud generation,

    S. Luo and W. Hu, “Diffusion probabilistic models for 3d point cloud generation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2021, Conference Proceedings, pp. 2837–2845

    2021

  37. [45]

    Repcd-net: Feature-aware recurrent point cloud denoising network,

    H. Chen, Z. Wei, X. Li, Y . Xu, M. Wei, and J. Wang, “Repcd-net: Feature-aware recurrent point cloud denoising network,” International Journal of Computer Vision , vol. 130, no. 3, pp. 615–629, 2022

    2022

  38. [46]

    Denoising point clouds in latent space via graph convolution and invertible neural network,

    A. Mao, B. Yan, Z. Ma, and Y . He, “Denoising point clouds in latent space via graph convolution and invertible neural network,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 5768–5777

    2024

  39. [47]

    Accelerating 3d deep learning with pytorch3d,

    N. Ravi, J. Reizenstein, D. Novotny, T. Gordon, W.-Y . Lo, J. Johnson, and G. Gkioxari, “Accelerating 3d deep learning with pytorch3d,” ArXiv,

  40. [48]

    Pu-net: Point cloud upsampling network,

    L. Yu, X. Li, C.-W. Fu, D. Cohen-Or, and P.-A. Heng, “Pu-net: Point cloud upsampling network,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018, Conference Proceedings

    2018

  41. [49]

    Paris-rue- madame database - a 3d mobile laser scanner dataset for benchmarking urban detection, segmentation and classification methods,

    A. Serna, B. Marcotegui, F. Goulette, and J.-E. Deschaud, “Paris-rue- madame database - a 3d mobile laser scanner dataset for benchmarking urban detection, segmentation and classification methods,” in ICPRAM, 2014

    2014

  42. [50]

    Mesh denoising via cascaded normal regression,

    P.-S. Wang, Y . Liu, and X. Tong, “Mesh denoising via cascaded normal regression,” ACM Trans. Graph., vol. 35, no. 6, p. Article 232, 2016. Dasith de Silva Edirimuni received the BSc degree in Physics from Hardin-Simmons University, Texas, USA in 2013 and the MSc degree in Phy...

    2016

  43. [2020]

    Available: https://arxiv.org/abs/2007.08501

    [Online]. Available: https://arxiv.org/abs/2007.08501

    2007 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.