Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Point Cloud Segmentation of Agricultural Vehicles using 3D Gaussian Splatting

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a point cloud segmentation model trained and validated exclusively on synthetic LiDAR data generated from 3D Gaussian Splatting meshes reaches a mean IoU of 91.35% on real-world agricultural vehicle point clouds.

desk verdict Useful synthetic-data pipeline for niche point-cloud segmentation, but the headline generalization result is undercut by possible instance-level leakage and unmatched training sizes. read the letter →

arxiv 2506.05009 v1 pith:CP2C55ZQ submitted 2025-06-05 cs.CV

Alfred T. Christiansen , Andreas H. H{o}jrup , Morten K. Stephansen , Md Ibtihaj A. Sakib , Taman S. Poojary , Filip Slezak , Morten S. Laursen , Thomas B. Moeslund
show 1 more author
Joakim B. Haurum
This is my paper · ORCID
classification cs.CV
keywords pointcloudsegmentation3DGaussianSplattingOpacityFieldssyntheticdataLiDARsimulationagriculturalvehiclessemanticGazebo
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

This paper proposes a pipeline for generating realistic synthetic LiDAR point clouds of agricultural vehicles without hand-built 3D models. The pipeline captures drone images of real tractors and combine harvesters, reconstructs them as meshes via 3D Gaussian Splatting and Gaussian Opacity Fields, and places the meshes in the Gazebo simulator with a simulated LiDAR. Training PointNet++, Point Transformer V3, and OACNN only on such synthetic clouds yields up to 91.35% mean IoU on a real 3,000-cloud test set, and performance improves when the synthetic dataset is enlarged. The authors argue that this makes synthetic data a viable path for niche domains where real annotated point clouds are scarce.

What carries the argument

The pipeline's central object is the mesh generation chain: drone-captured images are processed with COLMAP's structure-from-motion, turned into a 3D Gaussian Splatting scene, and converted to a mesh by Gaussian Opacity Fields, which directly extracts surface normals from the Gaussians and uses Marching Tetrahedra for the final mesh. These meshes, together with some grass and environment meshes, are loaded into Gazebo, where a custom plugin randomizes vehicle positions and a simulated Ouster OS0-128 LiDAR produces annotated point clouds. This chain carries the argument because it makes synthetic data that closely mimics real LiDAR sampling patterns while providing perfect per-point labels.

What would settle it

Hand-label a random subset of the 3,000 real test point clouds, or have human workers verify the cluster assignments, and recompute mean IoU for the synthetic-only PTv3 model; if the corrected labels shift the number by more than a few percentage points, the claimed synthetic-to-real transfer is inflated by label noise.

Watch

Extended reading notes

Core claim

The central claim is that high-fidelity meshes extracted from Gaussian Splatting reconstructions can replace hand-modeled assets in a LiDAR simulator well enough that segmentation models trained and validated exclusively on synthetic point clouds perform near or above real-data baselines on real test data. Point Transformer V3 trained on 65,000 synthetic clouds reports a mean IoU of 91.35% on 3,000 real point clouds, and synthetically trained PointNet++ outperforms the real-only baseline. The models also generalize to tractor models held out from training, with a mean IoU drop for unseen tractors of only 3.65 percentage points compared to seen tractors.

Load-bearing premise

The real-world test labels are generated automatically by KISS-ICP registration plus clustering, with no reported validation against manual annotation, so any error in those labels directly weakens every IoU comparison.

Editorial extensions

If this is right

  • Synthetic-only training can substitute for a substantial fraction of real annotated data in niche domains where real LiDAR labels are hard to obtain.
  • Scaling the synthetic dataset from 10,000 to 65,000 point clouds raises mean IoU for all three models; the 65k PTv3 and OACNN models exceed 91% and approach the real-data baselines.
  • Mixed real-plus-synthetic training improves mean IoU most when real data is scarce, with an average gain of 5.5 percentage points when only 20% of the available real data is used.
  • The method generalizes to unseen instances within a class; models trained without a specific tractor model lose only about 3.65 mean IoU points on that held-out tractor in real test data.
  • Because the simulator models the LiDAR geometrically, changing sensor specifications requires only reconfiguring the LiDAR plugin, not recollecting real data.

Reading between the lines

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

  • Because the synthetic pipeline yields perfect labels, the reported gains may partly reflect label quality rather than geometric realism; a study that corrupts synthetic labels to match real annotation noise would separate these factors.
  • The same pipeline could extend to other vehicle types or infrastructure objects if one can fly a drone around the object, with the main constraint being mesh extraction quality for thin or reflective structures.
  • The tall-grass failure observed in the qualitative analysis suggests the simulator's environment meshes are incomplete; adding procedurally generated vegetation to the Gazebo scene could be a direct next test.
  • The claim that synthetic-only training approaches real-data performance is tied to this specific test set and sensor; re-running the pipeline on a public dataset like SemanticKITTI would clarify how far the result generalizes.
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

4 major / 5 minor

Summary. The paper proposes a pipeline for generating synthetic LiDAR point cloud datasets for semantic segmentation of agricultural vehicles. The pipeline uses drone-captured images to reconstruct 3D meshes of vehicles via Gaussian Opacity Fields (GOF), places those meshes in a Gazebo simulation, and renders point clouds with a simulated Ouster OS0-128 LiDAR. Three segmentation models (PointNet++, PTv3, OACNN) are trained and validated exclusively on synthetic data and then evaluated on a real test set of 3,000 point clouds. The authors report that PTv3 reaches 91.35% mIoU with 65k synthetic point clouds, that other models also achieve high performance, and that leave-one-tractor-out experiments show a limited performance drop for unseen tractors. Mixed real+synthetic training results are presented in the supplementary material.

Significance. The pipeline is a potentially useful contribution for niche domains with scarce real data, as it combines 3D Gaussian Splatting mesh extraction with LiDAR simulation in a flexible and relatively low-cost way. The paper includes a real held-out test set, two synthetic dataset sizes, and a leave-one-tractor-out generalization test, which are stronger evaluation elements than typical synthetic-data papers. The experimental scope (three architectures, ablations on epochs/learning rate/downsampling) is solid. However, the central claim of sim-to-real generalization is currently confounded by a likely instance-level overlap between the reconstructed vehicles and the vehicles in the real test set, which must be resolved before the headline numbers can be interpreted as evidence of generalization to unseen objects.

major comments (4)
  1. [§3.1, §3.2, §5.3] The synthetic meshes are generated from drone captures of 'each vehicle' (Section 3.1), and the real LiDAR test set is described as captured on a farm 'with multiple tractors and combine harvesters' (Section 3.2), with the test split balanced 'across all possible tractor and combine models captured in the dataset.' If the drone-captured vehicles are the same physical instances as those in the LiDAR test set, the synthetic training data contains instance-level geometry of the test objects. The leave-one-tractor-out experiment (Section 5.2, Table 2) partially addresses this for tractors, but the headline 91.35% mIoU (abstract) uses all assets, and no equivalent hold-out is reported for combine harvesters. To support the claim of generalization to unseen objects, please (i) state explicitly whether the drone-scanned instances are disjoint from the LiDAR test instances, or (ii) provide a combined hold-out experiment for both tractors and combines, or (iii) re-frame the central claim as testing asset fidelity rather than sim-to-real generalization.
  2. [§3.2] The real-world ground truth labels are produced by KISS-ICP registration followed by clustering, with no quantitative validation of the resulting labels. Since every IoU and mIoU number in Tables 1, 2, and 5 is computed against these labels, systematic label noise (e.g., mis-assigned cluster boundaries, registration drift) would affect all comparisons. Please include a validation of the annotation quality, such as a comparison with manual labels on a subset of point clouds or an inter-annotator agreement measure, to establish the reliability of the evaluation reference.
  3. [§4.1, Supplementary A.1–A.2] The text states that a small search over learning rate and epochs was performed, but it does not specify which validation split was used for this search. The main protocol validates on synthetic data (Section 4.2), but this must be made explicit for the hyperparameter sweep as well. If the real test set was used for any hyperparameter selection, the reported performance is optimistic. Please clarify the evaluation set used in Tables 3 and 4.
  4. [§5.3, Table 1, Abstract] The abstract reports 91.35% mIoU for PTv3, but the main text only provides qualitative statements such as 'over 90%' in Figure 9, and Table 1 contains only the 10k synthetic results. Please add a table with exact IoU/mIoU values for the 65k synthetic dataset. In addition, all results appear to be single training runs with no error bars or multiple seeds; given that the comparison between synthetic-only and real-only models sometimes shows small differences (e.g., PointNet++), standard deviations across seeds are needed to assess whether these differences are meaningful.
minor comments (5)
  1. [§3.1] In Section 3.1, 'Figure??' should be replaced with the correct figure number (likely Figure 2, which shows the dataset acquisition pipeline).
  2. [Table 2 caption] The caption reads 'Figure shows the mean IoU of tractor models...' but it should read 'Table shows...'.
  3. [Abstract and §7] The claim that the models 'generalize across semantic classes' is not supported by the experiments, which test generalization to unseen tractor instances within the same semantic class, not across different semantic classes. Please reword to 'generalize to unseen instances of the same class' or provide a cross-class experiment.
  4. [§5.4, Figure 10] The qualitative example discusses a trailer misclassified as a combine harvester, but trailer is not one of the three annotated classes; it presumably falls under 'other.' Please clarify this in the text and figure caption to avoid confusion.
  5. [Supplementary A.2] In Table 4, the learning rate scaling factors are described, but the actual base learning rates after batch-size adjustment are not printed. Please report the exact learning rates used for each model in the sweep.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports empirical sim-to-real evaluations, with no fitted parameter or derivation that reduces to its inputs.

full rationale

The paper contains no derivation chain that reduces to its own inputs. The central claim is empirical: models trained and validated only on synthetic point clouds are tested on a fixed, external set of 3,000 real LiDAR point clouds. The synthetic data are generated independently of the segmentation models: drone images of vehicles are reconstructed into meshes via GOF, placed in Gazebo, and scanned with a simulated LiDAR. The models are not used in any way to generate the synthetic data, and no parameter is fitted to the real test set. Model selection is performed on synthetic validation data (Section 4.2: 'the models are trained on 10,000 synthetic point clouds and validated on 2,000 synthetic only point clouds'), and the real test set is only used once at evaluation time. The cited works (GOF, KISS-ICP, PTv3, OACNN, PointNet++) are external methods; the only reference with overlapping authors ([7], 'From neRF to 3DGS') is a related-work citation and is not load-bearing for any conclusion. The leave-one-tractor-out experiment (Section 4.3, Figure 7, Table 2) directly tests generalization to unseen instances, and although combine harvesters are not held out, this is a completeness limitation rather than circularity. The concern that drone-scanned vehicles may be the same instances as those in the real test set is a potential instance-level leakage confound that weakens the generality of the 91.35% mIoU number, but it is not circular reasoning: the synthetic training point clouds are not derived from the real test point clouds, and the reported numbers are not constructed from the training data by definition. Similarly, the reliance on KISS-ICP and clustering for ground-truth labels is an annotation-quality concern, not a circularity. Overall, the paper's evaluation is self-contained against an external benchmark, so the circularity score is 0.

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

The paper's empirical claims rest on unvalidated assumptions about annotation accuracy, mesh fidelity, and LiDAR simulation fidelity, plus hand-chosen hyperparameters and dataset sizes. No invented physical entities are introduced.

free parameters (4)
  • Training epochs = 20 (default; sweep over 10, 20, 30 in Table 3)
    Main results use 20 epochs; supplementary shows higher epochs improve mIoU, so the choice affects reported numbers but is a standard hyperparameter.
  • Learning rate multiplier = 1.0 (base rates 2e-3 for PointNet++/OACNN, 5e-3 for PTv3, scaled with batch size 32)
    From original implementations plus a small sweep; reported numbers would change with other multipliers (Table 4).
  • Point cloud downsampling limit = 30,000 (PointNet++), 40,000 (OACNN, PTv3)
    Chosen due to VRAM constraints; downsampling ablation shows performance is stable but not identical.
  • Synthetic dataset size = 10,000 and 65,000 point clouds
    Experimentally chosen sizes; the 65k set drives the improvement reported in Section 5.3.
assumptions (4)
  • domain assumption KISS-ICP registration and clustering yield accurate per-point labels for the real LiDAR sequences.
    Section 3.2 describes annotation but provides no quantitative validation; all evaluation IoU values are computed against these labels.
  • domain assumption The Gazebo LiDAR plugin faithfully simulates the geometric beam pattern of the Ouster OS0-128, and the meshes placed in the scene are geometrically accurate.
    Section 3.3 relies on this for sim-to-real transfer; no quantitative comparison between simulated and real point cloud geometry is given.
  • domain assumption The GOF-extracted meshes are sufficiently faithful representations of the seven tractor and three combine models.
    Section 3.3 compares GOF to SuGaR/2DGS qualitatively and chooses GOF; fidelity is only visually assessed.
  • domain assumption Semantic classes are limited to three (tractor, combine, other), and the test set distribution is representative.
    Section 3.2 defines classes; the extreme class imbalance (80.9% other) affects all mIoU numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Point Cloud Segmentation of Agricultural Vehicles using 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/CP2C55ZQ

@misc{pith2026250605009,
  author       = {Pith},
  title        = {Pith review of: Point Cloud Segmentation of Agricultural Vehicles using 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CP2C55ZQ}},
  note         = {Machine review of arXiv:2506.05009}
}
read the original abstract

Training neural networks for tasks such as 3D point cloud semantic segmentation demands extensive datasets, yet obtaining and annotating real-world point clouds is costly and labor-intensive. This work aims to introduce a novel pipeline for generating realistic synthetic data, by leveraging 3D Gaussian Splatting (3DGS) and Gaussian Opacity Fields (GOF) to generate 3D assets of multiple different agricultural vehicles instead of using generic models. These assets are placed in a simulated environment, where the point clouds are generated using a simulated LiDAR. This is a flexible approach that allows changing the LiDAR specifications without incurring additional costs. We evaluated the impact of synthetic data on segmentation models such as PointNet++, Point Transformer V3, and OACNN, by training and validating the models only on synthetic data. Remarkably, the PTv3 model had an mIoU of 91.35\%, a noteworthy result given that the model had neither been trained nor validated on any real data. Further studies even suggested that in certain scenarios the models trained only on synthetically generated data performed better than models trained on real-world data. Finally, experiments demonstrated that the models can generalize across semantic classes, enabling accurate predictions on mesh models they were never trained on.

Figures

Figures reproduced from arXiv: 2506.05009 by the authors.

Figure 1
Figure 1. Illustration shows the pipeline created for generating the synthetic point cloud dataset. Step (1) shows the process of gathering [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the pipeline used to obtain synthetic [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Drone-captured frames of a tractor (left) and a combine [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: The Gazebo simulation where the target assets have [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: An annotated point cloud obtained from the gazebo [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Figure shows the mIoU results for each test in which the [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: OACNN IoU of tractors when not in training set [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Figure shows the comparison between real only trained [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: We compare the point predictions of a OACNN network trained in the baseline configuration versus trained with the Synth [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 11
Figure 11. Figure 11: Figure shows the mIoU results for the OACNN model, [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Figure shows the mIoU results for the OACNN model, [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: other, tractor, and combine harvester Top row: tractor with large trailer, second row: tractor with small trailer, Third row tractor with tall grass backdrop [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages

  1. [1]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 3

    2022

  2. [2]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anto-aliased grid-based neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19697–19705, 2023. 3

    work page 2023

  3. [3]

    Se- mantickitti: A dataset for semantic scene understanding of lidar sequences, 2019

    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, 2019. 8

    work page 2019

  4. [4]

    Pointmixup: Augmentation for point clouds

    Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. Pointmixup: Augmentation for point clouds. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part III 16, pages 330–345. Springer, 2020. 1, 2

    work page 2020

  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 Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 3

    work page 2017

  6. [6]

    Carla: An open urban driv- ing simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Anto- nio Lopez, and Vladlen Koltun. Carla: An open urban driv- ing simulator. InConference on robot learning, pages 1–16. PMLR, 2017. 2

    work page 2017

  7. [7]

    Moeslund

    Magnus Kaufmann Gjerde, Filip Slezák, Joakim Bruslund Haurum, and Thomas B. Moeslund. From neRF to 3DGS: A leap in stereo dataset quality? InSynthetic Data for Com- puter Vision Workshop @ CVPR 2024, 2024. 3

    work page 2024

  8. [8]

    Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering

    Antoine Guédon and Vincent Lepetit. Sugar: Surface- aligned gaussian splatting for efficient 3d mesh reconstruc- tion and high-quality mesh rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354–5363, 2024. 3, 4

    work page 2024

Show all 39 references
  1. [9]

    Baking neural ra- diance fields for real-time view synthesis

    Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural ra- diance fields for real-time view synthesis. InProceedings of the IEEE/CVF international conference on computer vision, pages 5875–5884, 2021. 3

    2021

  2. [10]

    2d gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically ac- curate radiance fields. InACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3, 4

    2024

  3. [11]

    3d gaussian splatting for real-time radiance field rendering, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 3

    2023

  4. [12]

    Design and use paradigms for gazebo, an open-source multi-robot simula- tor

    Nathan Koenig and Andrew Howard. Design and use paradigms for gazebo, an open-source multi-robot simula- tor. In2004 IEEE/RSJ international conference on intelli- gent robots and systems (IROS)(IEEE Cat. No. 04CH37566), pages 2149–2154. Ieee, 2004. 3

    2004

  5. [13]

    Pointaugment: an auto-augmentation framework for point cloud classification

    Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. Pointaugment: an auto-augmentation framework for point cloud classification. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 6378–6387, 2020. 1, 2

    2020

  6. [14]

    Pointcnn: Convolution on x-transformed points.Advances in neural information processing systems, 31, 2018

    Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points.Advances in neural information processing systems, 31, 2018. 1, 2

    2018

  7. [15]

    Adfactory: An effective framework for gen- eralizing optical flow with nerf

    Han Ling, Quansen Sun, Yinghui Sun, Xian Xu, and Xingfeng Li. Adfactory: An effective framework for gen- eralizing optical flow with nerf. In2024 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 20591–20600, 2024. 3

    2024

  8. [16]

    Jong Won Ma, Thomas Czerniawski, and Fernanda Leite. Semantic segmentation of point clouds of building interiors with deep learning: Augmenting training datasets with syn- thetic bim-based point clouds.Automation in construction, 113:103144, 2020. 1, 2, 4

    2020

  9. [17]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.arXiv preprint arXiv:2003.08934, 2020

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.arXiv preprint arXiv:2003.08934, 2020. 3

    2003 arXiv

  10. [18]

    Ouster os0: High-precision ultra wide, 2024

    Ouster. Ouster os0: High-precision ultra wide, 2024. Ac- cessed on November 22, 2024. 3

    2024

  11. [19]

    Oa-cnns: Omni- adaptive sparse cnns for 3d semantic segmentation, 2024

    Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Heng- shuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni- adaptive sparse cnns for 3d semantic segmentation, 2024. 1, 2, 7

    2024

  12. [20]

    Qi, Hao Su, Kaichun Mo, and Leonidas J

    Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation, 2017. 2

    2017

  13. [21]

    Qi, Li Yi, Hao Su, and Leonidas J

    Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Point- net++: Deep hierarchical feature learning on point sets in a metric space, 2017. 1, 2, 7

    2017

  14. [22]

    Merf: Memory-efficient radiance fields for real- time view synthesis in unbounded scenes.ACM Transactions on Graphics (TOG), 42(4):1–12, 2023

    Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srini- vasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Pe- ter Hedman. Merf: Memory-efficient radiance fields for real- time view synthesis in unbounded scenes.ACM Transactions on Graphics (TOG), 42(4):1–12, 2023. 3

    2023

  15. [23]

    Self-evolving depth-supervised 3d gaussian splatting from rendered stereo pairs

    Sadra Safadoust, Fabio Tosi, Fatma Güney, and Matteo Poggi. Self-evolving depth-supervised 3d gaussian splatting from rendered stereo pairs. InBritish Machine Vision Con- ference (BMVC), 2024. 3

    2024

  16. [24]

    Structure-from-motion revisited

    Johannes Lutz Schönberger and Jan-Michael Frahm. Structure-from-motion revisited. InConference on Com- puter Vision and Pattern Recognition (CVPR), 2016. 4

    2016

  17. [25]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Et- tinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhan...

    2020

  18. [26]

    3d segmentation of humans in point clouds with syn- thetic data

    Ayça Takmaz, Jonas Schult, Irem Kaftan, Mertcan Akçay, Bastian Leibe, Robert Sumner, Francis Engelmann, and Siyu Tang. 3d segmentation of humans in point clouds with syn- thetic data. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1292–1304,

  19. [27]

    Nerf-supervised deep stereo

    Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Mat- teo Poggi. Nerf-supervised deep stereo. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 855–866, 2023. 3

    2023

  20. [28]

    Attention is all you need.Advances in Neural Information Processing Systems, 2017

    A Vaswani. Attention is all you need.Advances in Neural Information Processing Systems, 2017. 2

    2017

  21. [29]

    KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way.IEEE Robotics and Automation Letters (RA-L), 8(2):1029–1036, 2023

    Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyrill Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way.IEEE Robotics and Automation Letters (RA-L), 8(2):1029–1036, 2023. 4, 6

    2023

  22. [30]

    Auto- matic generation of synthetic lidar point clouds for 3-d data analysis.IEEE Transactions on Instrumentation and Mea- surement, 68(7):2671–2673, 2019

    Fei Wang, Yan Zhuang, Hong Gu, and Huosheng Hu. Auto- matic generation of synthetic lidar point clouds for 3-d data analysis.IEEE Transactions on Instrumentation and Mea- surement, 68(7):2671–2673, 2019. 1, 2

    2019

  23. [31]

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

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds.ACM Transactions on Graphics (tog), 38(5):1–12, 2019. 1, 2

    2019

  24. [32]

    Point transformer v2: Grouped vector atten- tion and partition-based pooling.Advances in Neural Infor- mation Processing Systems, 35:33330–33342, 2022

    Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Heng- shuang Zhao. Point transformer v2: Grouped vector atten- tion and partition-based pooling.Advances in Neural Infor- mation Processing Systems, 35:33330–33342, 2022. 2

    2022

  25. [33]

    Point transformer v3: Simpler, faster, stronger, 2024

    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, 2024. 1, 2, 7

    2024

  26. [34]

    Synlidar: Learning from synthetic lidar sequential point cloud for semantic segmentation.CoRR, abs/2107.05399, 2021

    Aoran Xiao, Jiaxing Huang, Dayan Guan, Fangneng Zhan, and Shijian Lu. Synlidar: Learning from synthetic lidar sequential point cloud for semantic segmentation.CoRR, abs/2107.05399, 2021. 2, 4

    2021 arXiv

  27. [35]

    Bakedsdf: Meshing neural sdfs for real- time view synthesis

    Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P Srinivasan, Richard Szeliski, Jonathan T Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real- time view synthesis. InACM SIGGRAPH 2023 Conference Proceedings, pages 1–9, 2023. 3

    2023

  28. [36]

    Gaussian opacity fields: Efficient and compact surface reconstruc- tion in unbounded scenes.arXiv preprint arXiv:2404.10772,

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruc- tion in unbounded scenes.arXiv preprint arXiv:2404.10772,

    arXiv

  29. [37]

    A lidar point cloud generator: from a virtual world to autonomous driving

    Xiangyu Yue, Bichen Wu, Sanjit A Seshia, Kurt Keutzer, and Alberto L Sangiovanni-Vincentelli. A lidar point cloud generator: from a virtual world to autonomous driving. In Proceedings of the 2018 ACM on international conference on multimedia retrieval, pages 458–464, 2018. 1, 2

    2018

  30. [38]

    Place: Proximity learning of articulation and con- tact in 3d environments

    Siwei Zhang, Yan Zhang, Qianli Ma, Michael J Black, and Siyu Tang. Place: Proximity learning of articulation and con- tact in 3d environments. In2020 International Conference on 3D Vision (3DV), pages 642–651. IEEE, 2020. 3

    2020

  31. [39]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 16259–16268, 2021. 1, 2 Point Cloud Segmentation of Agricultural V ehicles using 3D Gaussian Splatting ...

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.