Pith. sign in

REVIEW 3 major objections 5 minor 48 references

FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction

T0 review · 3 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 FastPoint claims a 2.55x geomean end-to-end speedup for PointNet++-style 3D point cloud models by replacing most farthest-point-sampling iterations with a predicted minimum-distance curve, with mIoU changes within about 0.1 points.

desk verdict Solid, novel acceleration for FPS and neighbor search in PointNet++-style models, with a measured 2.55x speedup at parity accuracy; the k-NN 'no approximation' claim rests on an unproven coverage assumption. read the letter →

arxiv 2507.23480 v1 pith:3GQZSB3D submitted 2025-07-31 cs.CV

Donghyun Lee , Dawoon Jeong , Jae W. Lee , Hongil Yoon This is my paper
classification cs.CV
keywords farthestpointsamplingcloudinferenceaccelerationdistancecurvepredictionneighborsearchreuseNet++modelsGPUapproximate
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

FastPoint is a software-only method for speeding up 3D point cloud neural networks whose cost is dominated by farthest point sampling (FPS) and neighbor search. It claims that the minimum-distance curve produced by FPS is predictable: after roughly the first tenth of the iterations, the distances between successive sampled points fall on a smooth decreasing curve. A small trained estimator predicts the rest of that curve, and sampling is driven by the predicted curve through exclusion lists that are built in one fully parallel pass. The same exclusion lists are then reused for ball query and to narrow $k$-NN search, avoiding redundant distance computations. Integrated into the first layer of PointVector and PointMetaBase models, the method reports a geomean 2.55x end-to-end speedup on an RTX 3090 GPU while keeping mIoU within about 0.1 points of the FPS baseline.

What carries the argument

The load-bearing object is the predicted minimum-distance curve $C(t)$ for FPS iterations, obtained by fitting a lightweight three-layer multilayer perceptron (MLP, 32-128-128-64) on the first $p=0.1$ fraction of the true curve and outputting the remaining 0.9. The curve is divided into segments; each segment has a radius threshold equal to the predicted minimum distance at the segment boundary, and an exclusion list stores all input points within that radius of each sampled point. Bitmaps over the exclusion lists drive sampling: after choosing a point, all points within its neighborhood for every remaining segment are marked unavailable, so candidate points are found by scanning a bitmap rather than by recomputing distances. The same exclusion structure is reused for ball query and as a search-space filter for $k$-NN, eliminating redundant distance computations.

What would settle it

Run FastPoint on a sparse, unevenly distributed point cloud such as a long-range LiDAR scan and compare the nearest neighbors returned by Redundancy-Free $k$-NN against exact $k$-NN; if any true neighbor is absent from the Segment-1 exclusion list, or if mIoU drops by more than 0.1 points on that scene, the no-accuracy-loss claim fails in that regime.

Watch

Extended reading notes

Core claim

The central claim is that farthest point sampling need not be computed exhaustively: the maximized minimum-distance value across iterations follows a smooth decreasing curve, and the first 10% of FPS iterations capture the curve's shape well enough to predict the remaining 90%. FastPoint therefore decouples sampling from distance computation. It uses a three-layer MLP estimator to forecast the curve, segments the curve into per-segment radius thresholds, builds exclusion lists of points within each segment's radius, and samples only points not excluded, with early termination falling back to exact FPS if no points remain. Because the exclusion lists contain distances that would otherwise be recomputed, FastPoint also reuses them for ball query and prunes the $k$-NN search space. The paper argues this produces sampling quality above 99% of FPS's average minimum distance, accuracy differences within about 0.1 mIoU on the tested indoor and outdoor datasets, and a geomean 2.55x end-to-end speedup.

Load-bearing premise

The whole accuracy-preserving story rests on the predicted minimum-distance curve matching the true FPS curve closely enough that the radius-based exclusion lists behave like real FPS neighborhoods; in particular, the paper does not prove that the true $k$ nearest downsampled neighbors always lie within the predicted radius used for $k$-NN pruning.

Editorial extensions

If this is right

  • The first-layer FPS and neighbor search, which together dominate PointNet++-style inference, can be replaced without retraining the network.
  • The speedup grows with point-cloud size: reported end-to-end speedups rise from about 1.3-1.6x at 16k points to over 3.4x at 96k points.
  • FastPoint composes with other FPS accelerators: replacing the initial 10% FPS with QuickFPS lifts the geomean speedup to 2.76x over baseline FPS.
  • The approach transfers beyond PointNet++-style models to other FPS-based architectures: applying it to Point Transformer yields a 2.16x end-to-end speedup with no mIoU loss on S3DIS.
  • Because neighbor search reuses the exclusion list without approximation, the accuracy impact is attributed almost entirely to the sampling approximation.

Reading between the lines

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

  • If the minimum-distance curve is as predictable as the paper reports, the same estimator may apply to related greedy farthest-point variants, including training-time sampling, though the paper only evaluates inference.
  • The exactness of the $k$-NN pruning is the part most likely to break on unseen geometry: in a sparse region, the true $k$ nearest downsampled neighbors may lie beyond the Segment-1 radius $R_1$, and then the reported 'no accuracy loss' would not hold even though the method still runs fast.
  • A direct test would measure nearest-neighbor recall of Redundancy-Free $k$-NN on sparsely distributed outdoor scans and compare it with exact $k$-NN; the paper does not report this metric.
  • The estimator is dataset-specific, so deployment on a new sensor or scene distribution may need a fresh estimator or a fallback; cross-dataset results in the appendix show indoor-to-indoor transfer works better than outdoor-to-indoor.
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. FastPoint proposes a software-only acceleration scheme for PointNet++-style point cloud models. The key idea is to predict the FPS minimum-distance curve with a small MLP, segment the predicted curve, build per-segment exclusion lists, and then sample points from the non-excluded set at segment-level granularity, with early termination handing the remaining iterations back to exact FPS. The same exclusion lists are reused to accelerate ball query and to prune the search space of k-NN. Evaluated on S3DIS, ScanNet, and SemanticKITTI with PointVector and PointMetaBase, the paper reports sampling quality above 98% of FPS, mIoU differences within 0.12 points, and a geomean 2.55x end-to-end speedup on an RTX 3090, with further gains when combined with QuickFPS.

Significance. If the central claims are correct, this is a practically useful software-only acceleration result: it targets real bottlenecks in widely used PointNet++-style models, provides reproducible artifacts (source code link in the paper), and includes unusually broad comparisons against QuickFPS, random sampling, grid sampling, Adjustable FPS, and EdgePC. The observation that FPS minimum-distance curves are predictable and that sampling can be decoupled from distance computation is genuinely interesting. However, the paper's "no approximation" statement for the k-NN optimization is not yet supported, and the accuracy comparisons lack run-to-run variance, so the no-accuracy-sacrifice conclusion requires additional evidence.

major comments (3)
  1. [§4.2, §5.3] The claim in §5.3 that Redundancy-Free Neighbor Search "does not introduce any approximation" is not supported by the description in §4.2. The k-NN search space is restricted to the Segment-1 exclusion list, whose entries are points within distance R1 of the query. R1 is introduced in §4.1 as a lower bound on the minimum distance between sampled points in Segment 1, not as a covering radius for the downsampled set; a query in a sparse region can therefore have true k nearest downsampled neighbors at distances greater than R1, and restricting the search to the exclusion list would drop them. The ablation in Table 4a ("All") combines MDPS with the neighbor-search optimizations, so it cannot isolate the accuracy impact of this pruning. The authors should either prove that R1 covers the k nearest downsampled neighbors for every query (stating the required assumptions), or explicitly treat the k-NN as approximate and report its isolated accuracy impact.
  2. [Algorithm 2, lines 41-46] The Early Termination path initializes the FPS distance matrix for the remainder FPS by checking only sampled points that appear in excl_list_1[i]. This is exact only if every sampled point that could reduce dists[i] is already in the Segment-1 exclusion list, which is the same coverage condition on R1 as in §4.2. Without a guarantee that R1 is a covering radius for the sampled set, a sampled point outside the exclusion list could be closer to P[i] than the current dists[i], and the transition to exact FPS would be incorrect. The paper should state and justify this condition, or modify the initialization to consider all sampled points.
  3. [Tables 2 and 3] The accuracy claims rest on mIoU differences of at most 0.12 points, but MDPS includes random choices (the seed point and the "findAnyOne" selection among available points), and no seeds, repeated runs, or standard deviations are reported. Without this information it is impossible to tell whether the reported differences reflect true accuracy preservation or run-to-run noise. Given that the paper's headline is "without sacrificing accuracy," the authors should report mean and variance over multiple runs, or fix and state the random seed for every experiment.
minor comments (5)
  1. [Algorithm 2, line 26] The segment transition rule "seg <- max(div(i, n/nseg), seg)" uses an undefined div operation and an unclear update order; integer division and the intended semantics should be specified explicitly.
  2. [§4.1, Figure 4] The ordering of R1, R2, and R3 is stated inconsistently: the text says "dist(P0, P1) < R3 < R2 < R1," while the earlier description of segment-boundary radii and the statement that "Ri serves as the lower bound of the minimum distance in Segment i" suggest the opposite ordering. This should be clarified.
  3. [Table 4a and Table 4b] The label "Sematic KITTI" is misspelled and should read "SemanticKITTI."
  4. [References [38] and [39]] References [38] and [39] are both assigned arXiv:2304.06906; the Swin3D++ entry appears to have the wrong identifier and should be corrected.
  5. [Appendix A.3] The main text says "polynomial functions" were tried as estimators, while Appendix A.3 describes power functions; the terminology should be made consistent.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the curve prediction is a supervised fit with held-out evaluation, and the speedup and accuracy claims are measured against baselines; the only same-author citation is comparison-only.

full rationale

FastPoint's central derivation is empirical rather than circular. The minimum-distance curve estimator is a 3-layer MLP (Section 4.1, Appendix A.3) trained on ground-truth FPS curves from the training split and evaluated by MAPE on a separate validation split (Table 5), with reported errors of 0.77-1.93%. The predicted curve then sets segment radii, so any mismatch would directly appear as estimator error and degraded sampling quality; these quantities are not forced to match by construction. The headline results are measured end-to-end against baseline FPS and QuickFPS on held-out validation scenes (Section 5.4, Figures 7-10), and accuracy preservation is reported as measured mIoU differences (Tables 1-2), not derived from the fitted curve. The only same-author citation, L-FPS [14], is used in Appendix A.4 for comparison and is not load-bearing. The closest concern is the claim in Sections 4.2 and 5.3 that Redundancy-Free k-NN 'does not introduce any approximation'; the paper does not prove that all true k nearest downsampled neighbors lie within the Segment-1 radius R1, so this is an unsupported soundness assertion rather than a circular reduction, and it does not raise the circularity score under the stated criteria.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method rests on an empirical predictability claim about FPS distance curves, a dataset-fitted MLP, and an unstated coverage assumption for the k-NN search-space reduction. The exclusion-list construction is an algorithmic device, not a new physical entity.

free parameters (4)
  • p (initial FPS fraction) = 0.1
    Ratio of initial FPS iterations used to predict the rest of the curve; chosen via ablation in Appendix B.6.
  • nseg (segment count) = 6
    Number of distance curve segments; chosen empirically where accuracy saturates (Section 5.5).
  • MLP estimator weights = trained per dataset
    3-layer MLP trained on minimum distance curves from the training split of each dataset (Appendix A.3).
  • exclusion list radii = derived from predicted curve
    Radii are the predicted minimum distances at segment boundaries; they depend on the fitted MLP and thus inherit its fitted nature.
assumptions (3)
  • domain assumption The FPS minimum distance curve is smooth, decreasing, and predictable from its first 10%
    Section 4 states this as an empirical observation; the entire method relies on it.
  • ad hoc to paper The true k nearest downsampled neighbors of each query point are within the segment-1 exclusion radius R1
    Section 4.2 reduces the k-NN search space to the exclusion list without proving this coverage; not stated as an assumption.
  • domain assumption Per-dataset MLP generalization: an estimator trained on the training split transfers to the validation split and to similar datasets
    Appendix B.8 shows cross-dataset transfer but with higher error for outdoor-to-indoor; the central claim depends on this empirical generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction." pith.science (2026). https://pith.science/paper/3GQZSB3D

@misc{pith2026250723480,
  author       = {Pith},
  title        = {Pith review of: FastPoint: Accelerating 3D Point Cloud Model Inference via Sample Point Distance Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GQZSB3D}},
  note         = {Machine review of arXiv:2507.23480}
}
read the original abstract

Deep neural networks have revolutionized 3D point cloud processing, yet efficiently handling large and irregular point clouds remains challenging. To tackle this problem, we introduce FastPoint, a novel software-based acceleration technique that leverages the predictable distance trend between sampled points during farthest point sampling. By predicting the distance curve, we can efficiently identify subsequent sample points without exhaustively computing all pairwise distances. Our proposal substantially accelerates farthest point sampling and neighbor search operations while preserving sampling quality and model performance. By integrating FastPoint into state-of-the-art 3D point cloud models, we achieve 2.55x end-to-end speedup on NVIDIA RTX 3090 GPU without sacrificing accuracy.

Figures

Figures reproduced from arXiv: 2507.23480 by the authors.

Figure 1
Figure 1. Overview of PointNet++ Based Model Architecture [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Latency Breakdown of PointNet++ Based Models [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Motivation of Minimum Distance Curve Estimation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Overall Flow of Minimum Distance Prediction Sampling [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Redundancy Free Neighbor Search ing 1/10 of the original FPS iterations, adds only a frac￾tion of the original FPS latency. 2. Exclusion List Construction: The required all-to-all dis￾tance calculations are fully parallelizable, resulting in minimal latency overhead. D…
Figure 7
Figure 7. Figure 7: End-to-end speedup of FastPoint and QuickFPS. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: demonstrates the speedup specific to the sam￾pling operation. Across the S3DIS [1], ScanNet [5], and S3DIS ScanNet SemanticKITTI GeoMean 0 2 4 6 Speedup 8.21 FPS QuickFPS MDPS MDPS + QuickFPS [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Speedup of Redundancy Free Neighbor Search. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Speedup-mIoU curve of various sampling methods. PV [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: Minimum spacing distribution of each sampling [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 45 canonical work pages

  1. [1]

    Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese

    Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioan- nis Brilakis, Martin Fischer, and Silvio Savarese. 3d seman- tic parsing of large-scale indoor spaces. In CVPR, 2016. 7, 13

    work page 2016

  2. [2]

    Se- mantickitti: A dataset for semantic scene understanding of lidar sequences

    Jens Behley, Martin Garbade, Andres Milioto, Jan Quen- zel, Sven Behnke, Cyrill Stachniss, and Juergen Gall. Se- mantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 7, 13

    work page 2019

  3. [3]

    Largekernel3d: Scaling up kernels in 3d sparse cnns

    Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 2

    work page 2023

  4. [4]

    4d spatio-temporal convnets: Minkowski convolutional neural networks

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2

    work page 2019

  5. [5]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 7, 13

    work page 2017

  6. [6]

    Pointvector: A vector representation in point cloud analysis

    Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: A vector representation in point cloud analysis. In CVPR, 2023. 1, 2, 7

    work page 2023

  7. [7]

    Embracing Single Stride 3D Object Detector with Sparse Transformer

    Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing Single Stride 3D Object Detector with Sparse Transformer. In CVPR, 2022. 1, 2

    work page 2022

  8. [8]

    Mesorasi: Architecture support for point cloud analytics via delayed-aggregation

    Yu Feng, Boyuan Tian, Tiancheng Xu, Paul Whatmough, and Yuhao Zhu. Mesorasi: Architecture support for point cloud analytics via delayed-aggregation. In Proceedings of the 53th International Symposium on Microarchitecture (MI- CRO), 2020. 1

    work page 2020

Show all 48 references
  1. [9]

    Crescent: Taming memory irregularities for accelerating deep point cloud analytics

    Yu Feng, Gunnar Hammonds, Yiming Gan, and Yuhao Zhu. Crescent: Taming memory irregularities for accelerating deep point cloud analytics. In Proceedings of the 49th Annual International Symposium on Computer Architecture (ISCA), 2022. 1

    2022

  2. [10]

    Occuseg: Occupancy-aware 3d instance segmentation

    Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. InCVPR, 2020. 2

    2020

  3. [11]

    Quickfps

    Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chen- hao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps. http://github.com/hanm2019/bucket-based_ farthest-point-sampling_GPU. 2, 7, 13

  4. [12]

    Meng Han, Liang Wang, Limin Xiao, Hao Zhang, Chenhao Zhang, Xiangrong Xu, and Jianfeng Zhu. Quickfps: Archi- tecture and algorithm co-design for farthest point sampling in large-scale point clouds.IEEE Transactions on Computer- Aided Design of Integrated Circuits and Systems, ...

    2023

  5. [13]

    Randla-net: Efficient semantic segmentation of large-scale point clouds

    Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 2

    2020

  6. [14]

    Lee, and Hongil Yoon

    Donghyun Lee, Yejin Lee, Jae W. Lee, and Hongil Yoon. Frugal 3d point cloud model training via progressive near point filtering and fused aggregation. In ECCV, 2024. 3, 12

    2024

  7. [15]

    An adjustable farthest point sampling method for approximately-sorted point cloud data

    Jingtao Li, Jian Zhou, Yan Xiong, Xing Chen, and Chaitali Chakrabarti. An adjustable farthest point sampling method for approximately-sorted point cloud data. In 2022 IEEE Workshop on Signal Processing Systems (SiPS), 2022. 1, 3, 14

    2022

  8. [16]

    A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps

    Yaoxiu Lian, Xinhao Yang, Ke Hong, Yu Wang, Guohao Dai, and Ningyi Xu. A point transformer accelerator with fine-grained pipelines and distribution-aware dynamic fps. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1, 2

    2023

  9. [17]

    Pointmamba: A simple state space model for point cloud analysis

    Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaoqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In NeurIPS, 2024. 15

    2024

  10. [18]

    Meta architecure for point cloud analysis

    Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecure for point cloud analysis. In CVPR, 2023. 1, 2, 3, 7

    2023

  11. [19]

    Fusionarch: A fusion-based accelerator for point-based point cloud neural networks

    Xueyuan Liu, Zhuoran Song, Guohao Dai, Gang Li, Can Xiao, Yan Xiang, Dehui Kong, Ke Xu, and Xiaoyao Liang. Fusionarch: A fusion-based accelerator for point-based point cloud neural networks. In ACM/IEEE Design, Automation and Test in Europe Conference (DATE), 2024. 1

    2024

  12. [20]

    Densepoint: Learning densely contextual representation for efficient point cloud process- ing

    Yongcheng Liu, Bin Fan, Gaofeng Meng, Jiwen Lu, Shiming Xiang, and Chunhong Pan. Densepoint: Learning densely contextual representation for efficient point cloud process- ing. In ICCV, 2019. 2

    2019

  13. [21]

    Flatformer: Flattened window attention for effi- cient point cloud transformer

    Zhijian Liu, Xinyu Yang, Haotian Tang, Shang Yang, and Song Han. Flatformer: Flattened window attention for effi- cient point cloud transformer. In CVPR, 2023. 1

    2023

  14. [22]

    V oxel transformer for 3d object detection

    Jiageng Mao, Yujing Xue, Minzhe Niu, Haoyue Bai, Jiashi Feng, Xiaodan Liang, Hang Xu, and Chunjing Xu. V oxel transformer for 3d object detection. In ICCV, 2021. 2

    2021

  15. [23]

    Mix3D: Out-of-Context Data Aug- mentation for 3D Scenes

    Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann. Mix3D: Out-of-Context Data Aug- mentation for 3D Scenes. In International Conference on 3D Vision (3DV), 2021. 2

    2021

  16. [24]

    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. arXiv preprint arXiv:1612.00593, 2016. 1, 2 9

    2016 arXiv

  17. [25]

    Point- net++: Deep hierarchical feature learning on point sets in a metric space

    Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point- net++: Deep hierarchical feature learning on point sets in a metric space. arXiv preprint arXiv:1706.02413, 2017. 1, 2

    2017 arXiv

  18. [26]

    Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J

    Charles R. Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J. Guibas. Frustum pointnets for 3d object detec- tion from rgb-d data. In CVPR, 2018

    2018

  19. [27]

    Assanet: An anisotropical separable set abstraction for efficient point cloud representation learn- ing

    Guocheng Qian, Hasan Hammoud, Guohao Li, Ali Thabet, and Bernard Ghanem. Assanet: An anisotropical separable set abstraction for efficient point cloud representation learn- ing. In NeurIPS, 2021. 1

    2021

  20. [28]

    Pointnext: Revisiting pointnet++ with improved training and scaling strategies

    Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2

    2022

  21. [29]

    Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection

    Danila Rukhovich, Anna V orontsova, and Anton Konushin. Imvoxelnet: Image to voxels projection for monocular and multi-view general-purpose 3d object detection. In Proceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2

    2022

  22. [30]

    Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J

    Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2, 7

    2019

  23. [31]

    Sarma, Michael M

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 2019. 1, 2

    2019

  24. [32]

    Pointconv: Deep convolutional networks on 3d point clouds

    Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2

    2019

  25. [33]

    Point transformer v2: Grouped vector atten- tion and partition-based pooling

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

    2022

  26. [34]

    Point transformer v3: Simpler, faster, stronger

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

    2024

  27. [35]

    Paconv: Position adaptive convolution with dynamic ker- nel assembling on point clouds

    Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic ker- nel assembling on point clouds. In CVPR, 2021. 2

    2021

  28. [36]

    Grid-gcn for fast and scalable point cloud learning, 2020

    Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gcn for fast and scalable point cloud learning, 2020. 2

    2020

  29. [37]

    An efficient accelerator for point-based and voxel-based point cloud neural networks

    Xinhao Yang, Tianyu Fu, Guohao Dai, Shulin Zeng, Kai Zhong, Ke Hong, and Yu Wang. An efficient accelerator for point-based and voxel-based point cloud neural networks. In ACM/IEEE Design Automation Conference (DAC), 2023. 1, 2

    2023

  30. [39]

    Swin3d++: Effec- tive multi-source pretraining for 3d indoor scene understand- ing

    Yu-Qi Yang, Yu-Xiao Guo, and Yang Liu. Swin3d++: Effec- tive multi-source pretraining for 3d indoor scene understand- ing. arXiv preprint arXiv:2304.06906, 2024. 2

    2024 arXiv

  31. [40]

    Kandemir, and Chita R

    Ziyu Ying, Sandeepa Bhuyan, Yan Kang, Yingtian Zhang, Mahmut T. Kandemir, and Chita R. Das. Edgepc: Efficient deep learning analytics for point clouds on edge devices. In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023. 1, 2, 14

    2023

  32. [41]

    Efficient sampling and grouping acceleration for point cloud deep learning via sin- gle coordinate comparison

    Hyunsung Yoon and Jae-Joon Kim. Efficient sampling and grouping acceleration for point cloud deep learning via sin- gle coordinate comparison. In IEEE/ACM International Conference on Computer-Aided Design (ICCAD), 2023. 1

    2023

  33. [42]

    Point cloud mamba: Point cloud learning via state space model

    Tao Zhang, Haobo Yuan, Lu Qi, Jiangning Zhang, Qianyu Zhou, Shunping Ji, Shuicheng Yan, and Xiangtai Li. Point cloud mamba: Point cloud learning via state space model. In AAAI, 2025. 15

    2025

  34. [43]

    PointWeb: Enhancing local neighborhood features for point cloud processing

    Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. PointWeb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 2

    2019

  35. [44]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 15

    2021

  36. [45]

    Cylindrical and asymmetrical 3d convolution networks for lidar segmenta- tion

    Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmenta- tion. arXiv preprint arXiv:2011.10033, 2020. 1, 2 10 A. Supplementary Materials for MDPS A.1. MDPS Algorithm I...

    2011 arXiv

  37. [46]

    The key difference be- tween original FPS and this operation (Line 3-11) is that the maximum of minimum distance (i.e., max(dists)) must be saved at each iteration

    Minimum Distance Curve Estimation Minimum dis- tance curve estimation starts by performing FPS for 1/10 of the original number of iterations. The key difference be- tween original FPS and this operation (Line 3-11) is that the maximum of minimum distance (i.e., max(dists)) mus...

  38. [47]

    Distance Curve Segmentation We divide the distance curve into nseg segments and find the points within a speci- fied radius of segment boundaries (i.e., mdc[n ∗ seg/nseg]) for each input point. Distance between an input point P [i] and a query pointP [j] is calculated (Line 18...

  39. [48]

    Sampling consists of three main stages: bitmap update, sampling, and sampling availability check

    Sampling with Predicted Distance After initializing the bitmap for all segments to 1 (Line 24), sampling begins. Sampling consists of three main stages: bitmap update, sampling, and sampling availability check. First, we check the entry of exclusion list that corresponds to th...

  40. [49]

    For this tran- sition, the FPS distance matrix is initialized with the min- imum distances between the input points and the already sampled point set

    Early Termination If the sampling stage terminates before acquiring the desired number of points n, we make a transition to Farthest Point Sampling (FPS). For this tran- sition, the FPS distance matrix is initialized with the min- imum distances between the input points and th...

Pith tools

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