Pith. sign in

REVIEW 3 major objections 8 minor 73 references

Graph-Guided Dual-Level Augmentation for 3D Scene Segmentation

T0 review · 3 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that graph-guided dual-level augmentation—local geometric/semantic constraints plus global topological alignment—consistently improves 3D point cloud segmentation, lifting PTv3 to 79.8 mIoU on ScanNet, 75.5 on S3DIS, and…

desk verdict The local graph-guided augmentation is plausible and evaluated reasonably, but the global GGCL loss is non-differentiable as written and cannot confer the claimed gains. read the letter →

arxiv 2507.22668 v1 pith:T7HIZXWI submitted 2025-07-30 cs.CV

Hongbin Lin , Yifan Jiang , Juangui Xu , Jesse Jiaxi Xu , Yi Lu , Zhengyu Hu , Ying-Cong Chen , Hao Wang This is my paper · ORCID
classification cs.CV
keywords 3Dscenesegmentationpointcloudaugmentationobjectrelationshipgraphdual-levelconstraintsgraph-guidedrecompositionglobaltopologyindoorandoutdoorLiDARsemanticsynthesis
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 tries to show that data augmentation for 3D point cloud segmentation fails when it only perturbs local geometry or blindly recombines objects, because real scenes obey global spatial structure. It proposes a graph-guided framework that learns object co-occurrence and relationship statistics from training scenes, samples a target Object Relationship Graph (ORG), then places extracted objects under local geometric and semantic constraints and a global graph-alignment loss. The claim is that the resulting synthetic scenes are both diverse and realistic, and that adding them to training consistently improves segmentation across models and datasets. On the strongest baseline the reported gains are 79.8 versus 78.6 mIoU on ScanNet, 75.5 versus 74.7 on S3DIS, and 73.2 versus 72.3 on SemanticKITTI.

What carries the argument

The load-bearing object is the Object Relationship Graph (ORG), a stochastic graph whose node activation follows per-category Gaussian sampling regularized by Jensen-Shannon divergence to match the training-set class distribution, and whose edges are drawn from empirical co-occurrence probabilities. Seven hand-defined spatial relations give each edge a meaning and a corresponding loss. The global mechanism is the Graph Global Constraint Loss (GGCL), which combines node insertion and deletion counts, a minimum-cost node substitution over matchings, and the Frobenius difference of adjacency matrices between target and current graphs; its gradient is backpropagated into the 5-DOF pose parameters while a pretrained graph neural network is held fixed. This lets the optimization steer the whole layout toward the target topology rather than only fixing local overlaps.

What would settle it

Compute dL_topology/dθ for a representative generated scene by finite differences; if the numerical Jacobian disagrees with the paper's backprop, or is zero wherever an edge indicator or the argmin in Eq. (14) is active, then the global constraint is not optimizing poses as claimed. Ablating away the GGCL gradient while keeping the loss value would then be expected to leave layout and mIoU unchanged.

Watch

Extended reading notes

Core claim

The central claim is that explicitly enforcing the global topology of a scene—not just pairwise object relations—is what makes augmented 3D scenes useful for training segmenters. The paper models each scene as an Object Relationship Graph whose nodes are furniture categories and whose edges are seven spatial relationships (supported by, attached to, left of, right of, nearby, faces, oriented with), learned statistically from the training set. Generation samples a target graph from these statistics, then optimizes object poses to minimize collision and alignment losses, pairwise semantic relation losses, and a Graph Global Constraint Loss that aligns the generated graph's GNN embedding and adjacency with the target. The paper reports that this dual-level optimization outperforms existing augmentation, including Mix3D which PTv3 already uses, on indoor and outdoor benchmarks, and that removing the global term or replacing constraints with random insertion degrades performance.

Load-bearing premise

The global constraint only works if gradients can flow from the graph-alignment loss, which counts node insertions and deletions and takes a minimum over discrete matchings, into continuous object poses; the paper never specifies how this is done, and if those gradients vanish or are wrong the global term cannot actually steer layouts.

Editorial extensions

If this is right

  • Adding a 25 percent mix of graph-guided synthetic scenes improves mIoU for both OctFormer and PTv3 on all three benchmarks, including over PTv3's built-in Mix3D augmentation.
  • Constraint-free random insertion hurts performance (57.2 mIoU on S3DIS versus 75.5 for the full model), showing that structural plausibility is what makes the augmentation useful.
  • Augmentation scale matters: 10 percent already helps, 25 percent is best, and 50 percent drops performance, indicating a distributional-shift ceiling.
  • GT Sampling, which raises activation probability for the five worst-segmented classes, improves categories like picture, refrigerator, and other furniture with little cost to other classes.
  • The same pipeline transfers to outdoor LiDAR scenes and large-scale aerial data (STPLS3D), suggesting the graph priors generalize beyond indoor furniture.

Reading between the lines

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

  • A relaxed version that treats hard threshold edges as soft relations could turn the same idea into a controllable generative model for scene editing, not just segmentation training.
  • Because the ORG is learned per dataset, the framework implicitly claims that scene topology is transferable; one could test whether an ORG learned on ScanNet rooms produces plausible layouts for unseen room types or other indoor datasets without retraining.
  • The 50 percent augmentation degradation suggests the synthetic distribution drifts from the real one; a curriculum that anneals the mixing ratio or weights synthetic samples could push the optimal ratio higher.
  • Diversity is claimed but not directly measured; a graph-edit-distance or layout-novelty metric over generated ORGs would make the diversity claim testable against competing augmentations.
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 / 8 minor

Summary. The paper proposes a graph-guided data augmentation framework for 3D point cloud segmentation. It decomposes scenes into background and foreground objects, builds an Object Relationship Graph (ORG) from co-occurrence and spatial-relationship statistics, and synthesizes new scenes by placing sampled objects and refining their 5-DOF poses. The refinement jointly optimizes local geometric constraints (collision and surface alignment), local semantic constraints (seven spatial-relationship losses), and a global Graph Global Constraint Loss (GGCL) that compares the generated scene graph with the target ORG. Experiments with PTv3 and OctFormer on ScanNet, S3DIS, SemanticKITTI, and STPLS3D report consistent mIoU improvements, with the best result being PTv3+Ours at 79.8 mIoU on ScanNet versus 78.6 for the PTv3 baseline. The paper also includes ablations showing that the full local+global model outperforms local-only constraints on S3DIS and SemanticKITTI.

Significance. If the technical machinery is executable as claimed, the framework is a plausible contribution to 3D augmentation: it is evaluated on four datasets, covers both indoor and outdoor scenes, integrates with strong transformer baselines, and the authors commit to releasing code. The main scientific claim—consistent gains from graph-guided augmentation—is not circular, since the principal evaluation is against held-out test splits. However, the central novelty, the global GGCL optimization, is not specified in a way that can be executed as written, and the reported gains are small and are presented without variance estimates. Resolving the optimization specification and re-reporting the GT-sampling results on a clean split are necessary before the contribution can be assessed reliably.

major comments (3)
  1. [§3.5, Eq. (14)] The GGCL is not differentiable with respect to the 5-DOF pose parameters, yet §3.5 states that its gradient is backpropagated to adjust (x_i, y_i, z_i, theta_i, phi_i). The loss in Eq. (14) contains N_ins and N_del (integer counts), a minimum over node matchings pi in Pi, and ||A_target - A_current||_F, where the adjacency entries are activated by hard thresholds in Eq. (3) and Table 1. Each of these terms is piecewise constant in the poses, so the gradient is zero almost everywhere. No straight-through estimator, Sinkhorn/Gumbel relaxation, or continuous edge-activation surrogate is described in §3.5 or the appendix. Since the configuration 'Local + Global Constraints' is the best in Table 3 and the global constraint is the stated novelty, this is a load-bearing gap. Please provide a concrete relaxation or clarify that the layout refinement is a derivative-free search; without this, the ablation improvement attributed to the global constraint is unexplained.
  2. [Appendix A.6, §4.2, Table 9] The GT-sampling policy is selected from the validation performance of the baseline model: the five worst mIoU categories on ScanNet are chosen, their activation probabilities are tripled, and Table 9 then reports the improvement on that same setting. If the main ScanNet results include GT sampling, the evaluation is partly confounded by validation-set tuning. Please either report results on a held-out test split, or show that the category selection was made using a separate validation set and frozen before any reported test evaluation.
  3. [§4.2, Tables 2–3] The reported improvements are small in absolute terms (e.g., 0.4 and 0.3 mIoU in Table 3, and 1.2, 0.8, 0.9 mIoU in Table 2), and no error bars, repeated-seed statistics, or significance tests are provided. Given that the baselines are strong and already include sophisticated augmentations such as Mix3D, the central claim of 'consistent improvements' would be more convincing with at least a few repeated runs and a variance estimate, or a paired comparison showing that the gains are not within run-to-run noise.
minor comments (8)
  1. [§3.2] The phrase 'ruled-based Object Relationship Graph' should read 'rule-based Object Relationship Graph'.
  2. [Figure 2 caption] The caption uses 'Constrains' where 'Constraints' is intended.
  3. [§3.5, Eq. (14)] The symbols z_target and z_current are used without a definition; please state whether they are node feature matrices, graph embeddings, or node sets, and define how N_ins and N_del are computed from them.
  4. [§3.3, Eq. (4)] The collision loss also uses an indicator 1(BB_i ∩ BB_j ≠ ∅) and is therefore piecewise constant in the poses; if a continuous collision penalty is used in practice, it should be stated explicitly.
  5. [§3.6, Eq. (16)] The optimization procedure is underspecified: no optimizer, learning rate, iteration budget, or convergence criterion is given for the pose refinement. Please add these details or point to the corresponding code location.
  6. [Table 8] The row 'Without Poisson 79.49 Overall mIoU 79.79' appears inconsistent; the overall mIoU with and without Poisson should be reported consistently in the same table structure.
  7. [Table 4] The note for the 'Ours Only' column is confusing; clarify which model is evaluated on the generated data and what the column is meant to show.
  8. [Appendix A.6] There is a typo in the caption and text: 'refridgerator' should be 'refrigerator'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central augmentation claim is validated by held-out test accuracy and controlled ablations, not by its own construction.

full rationale

I walked the paper's derivation chain from ORG construction (Sec. 3.2) through the local and global constraints (Secs. 3.3-3.5) to the reported mIoU results (Tables 2-4). The guiding graph statistics are estimated from training data, and the augmented scenes are optimized to match those statistics, but the downstream claim is not equivalent to that construction: segmentation accuracy is measured on held-out ScanNet, S3DIS, and SemanticKITTI splits, and the ablation in Table 3 shows that random augmentation degrades performance while the full constraint model improves it, so the reported gains are not forced by the loss definitions. GT sampling in Appendix A.6 tunes activation probabilities to validation performance, which is benchmark tuning rather than circular reasoning. The non-differentiability of Eq. (14) noted in the reader's take is a correctness and implementability concern, not a circularity, because the paper does not define the loss in terms of the test metric. No load-bearing self-citation or imported uniqueness theorem is used; the only self-citations ([15,16]) are background references in the introduction. No circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on hand-picked thresholds and loss weights that are not reported, and on several domain assumptions about the sufficiency and transferability of learned relationship statistics. No new physical entities are introduced.

free parameters (5)
  • Loss weights (lambda_geo, lambda_sem, lambda_topo) = Not reported
    Weights in Eq. (15) balancing geometric, semantic, and topological losses. No sensitivity analysis is given.
  • Spatial relationship thresholds (tau, epsilon, tau_att, tau_dir, tau_left, tau_right, tau_face, epsilon'') = Not reported
    Thresholds in Table 1 defining support, attachment, direction, proximity, and orientation relationships. Values are not disclosed.
  • Edge activation threshold tau (Eq. 3) = Not reported
    Threshold controlling graph density and diversity during edge sampling.
  • GT sampling multiplier and category set = 3x, five worst classes from validation
    Category activation probabilities are tripled for categories selected from baseline validation results (Appendix A.6), a post-hoc choice.
  • Augmentation ratio = 25% (with 10% and 50% tested)
    The amount of generated data mixed into training; the paper selects 25% after testing other values, a hyperparameter choice.
assumptions (4)
  • domain assumption Seven hand-defined spatial relationships capture the structural regularities needed for augmentation.
    The entire generation pipeline relies on these relations (Table 1) to guide placement; missing relations would yield unrealistic scenes.
  • domain assumption Training-set co-occurrence and relationship statistics transfer to test scenes.
    The ORG is built from training data (Sec. 3.2); performance gains assume these statistics are representative of the test distribution.
  • domain assumption SceneGraphNet embeddings provide a meaningful topological distance between scene graphs.
    GGCL in Eq. (14) compares embeddings from a pretrained SceneGraphNet; if embeddings lack semantic structure, the global constraint is uninformative.
  • ad hoc to paper Gradient descent on pose parameters can minimize the GGCL despite hard relationship indicators and graph-edit counts.
    Sec. 3.5 claims backpropagation through L_topology, but the discrete operations in Table 1 and Eq. (14) are non-differentiable; no surrogate is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-Guided Dual-Level Augmentation for 3D Scene Segmentation." pith.science (2026). https://pith.science/paper/T7HIZXWI

@misc{pith2026250722668,
  author       = {Pith},
  title        = {Pith review of: Graph-Guided Dual-Level Augmentation for 3D Scene Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7HIZXWI}},
  note         = {Machine review of arXiv:2507.22668}
}
read the original abstract

3D point cloud segmentation aims to assign semantic labels to individual points in a scene for fine-grained spatial understanding. Existing methods typically adopt data augmentation to alleviate the burden of large-scale annotation. However, most augmentation strategies only focus on local transformations or semantic recomposition, lacking the consideration of global structural dependencies within scenes. To address this limitation, we propose a graph-guided data augmentation framework with dual-level constraints for realistic 3D scene synthesis. Our method learns object relationship statistics from real-world data to construct guiding graphs for scene generation. Local-level constraints enforce geometric plausibility and semantic consistency between objects, while global-level constraints maintain the topological structure of the scene by aligning the generated layout with the guiding graph. Extensive experiments on indoor and outdoor datasets demonstrate that our framework generates diverse and high-quality augmented scenes, leading to consistent improvements in point cloud segmentation performance across various models.

Figures

Figures reproduced from arXiv: 2507.22668 by the authors.

Figure 1
Figure 1. Single-Level Augmentation vs. Dual-Level Augmen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed dual-level point cloud data augmentation framework. The pipeline consists of three [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of Semantic Rules for Object Placement [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Visualization of augmented scenes generated by our method vs. Mix3D on S3DIS and SemanticKITTI datasets. The [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: t-SNE visualization of features extracted from the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualization of augmented scenes generated by our method on ScanNet dataset. The synthesized scenes are guided by [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Visualization of augmented scenes generated by our method on STPLS3D dataset. As there are more than 300 nodes in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 10
Figure 10. Figure 10: Comparison of mloU(%) with and without Poisson [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 8
Figure 8. Figure 8: An example of an object with none relationships to all surrounding furniture. The black point cloud in the figure represents a refrigerator. Due to incomplete scanning, only the front door of the refrigerator is captured. Such in￾complete objects, which exhibit only a …
Figure 9
Figure 9. Figure 9: Complete the boundary in Scannet a smooth surface 𝑆 that passes through these points. Mathemat￾ically, considering a signed distance function 𝑓 (𝑥) with 𝑥 ∈ R 3 , the surface 𝑆 is defined as the zero-level set of 𝑓 (𝑥). The Poisson equation for surface reconstruction i…
Figure 11
Figure 11. Figure 11: Comparison of mloU(%) Before and After GT Sam [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 64 canonical work pages

  1. [1]

    Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. 2018. Learning representations and generative models for 3d point clouds. (2018), 40–49

    work page 2018

  2. [2]

    Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 2016. 3d semantic parsing of large-scale indoor spaces. (2016), 1534–1543

    work page 2016

  3. [3]

    Behley, M

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall

    work page

  4. [4]

    Meida Chen, Qingyong Hu, Zifan Yu, Hugues Thomas, Andrew Feng, Yu Hou, Kyle McCullough, Fengbo Ren, and Lucio Soibelman. 2022. STPLS3D: A Large- Scale Synthetic and Real Aerial Photogrammetry 3D Point Cloud Dataset. In33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022

    work page 2022

  5. [5]

    Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. 2020. Pointmixup: Augmentation for point clouds. (2020), 330–345

    work page 2020

  6. [6]

    Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. 2023. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 21674–21683

    work page 2023

  7. [7]

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 2019. 4D Spatio- Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 3075–3084

    work page 2019

  8. [8]

    Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. 2020. Salsanext: Fast, uncertainty-aware semantic segmentation of lidar point clouds. In Advances in Visual Computing: 15th International Symposium, ISVC 2020, San Diego, CA, USA, October 5–7, 2020, Proceedings, Part II 15 . Springer, 207–222

    work page 2020

Show all 73 references
  1. [9]

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. 2017. Scannet: Richly-annotated 3d reconstructions of indoor scenes. (2017), 5828–5839

    2017

  2. [10]

    Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas Funkhouser, and Pat Hanrahan. 2012. Example-based synthesis of 3D object arrangements. ACM Transactions on Graphics (TOG) 31, 6 (2012), 1–11

    2012

  3. [11]

    Gege Gao, Weiyang Liu, Anpei Chen, Andreas Geiger, and Bernhard Schölkopf

  4. [12]

    Jingyu Gong, Zhou Ye, and Lizhuang Ma. 2022. Neighborhood co-occurrence modeling in 3D point cloud segmentation. Computational Visual Media 8 (2022), 303–315

    2022

  5. [13]

    Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. 2020. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence 43, 12 (2020), 4338–4364

    2020

  6. [14]

    Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. 2020. Randla-net: Efficient semantic segmen- tation of large-scale point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 11108–11117

    2020

  7. [15]

    Zhengyu Hu, Yichuan Li, Zhengyu Chen, Jingang Wang, Han Liu, Kyumin Lee, and Kaize Ding. 2024. Let’s Ask GNN: Empowering Large Language Model for Graph In-Context Learning. arXiv preprint arXiv:2410.07074 (2024)

    2024 arXiv

  8. [16]

    Zhengyu Hu, Jieyu Zhang, Haonan Wang, Siwei Liu, and Shangsong Liang. 2023. Leveraging relational graph neural network for transductive model ensemble. In Proceedings of the 29th ACM SIGKDD Conference on knowledge discovery and data mining. 775–787

    2023

  9. [17]

    Lutao Jiang, Hangyu Li, and Lin Wang. 2024. A General Framework to Boost 3D GS Initialization for Text-to-3D Generation by Lexical Richness. In Proceedings of the 32nd ACM International Conference on Multimedia . 6803–6812

    2024

  10. [18]

    Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. 2024. Diffindscene: Diffusion-based high-quality 3d indoor scene generation. (2024), 4526–4535

    2024

  11. [19]

    Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. 2006. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, Vol. 7

    2006

  12. [20]

    Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fa- had Shahbaz Khan, and Mubarak Shah. 2022. Transformers in vision: A survey. ACM computing surveys (CSUR) 54, 10s (2022), 1–41

    2022

  13. [21]

    Sihyeon Kim, Sanghyeok Lee, Dasol Hwang, Jaewon Lee, Seong Jae Hwang, and Hyunwoo J Kim. 2021. Point cloud augmentation with weighted local trans- formations. In Proceedings of the IEEE/CVF international conference on computer vision. 548–557

    2021

  14. [22]

    Thomas H Kolbe and Andreas Donaubauer. 2021. Semantic 3D city modeling and BIM. Urban informatics (2021), 609–636

    2021

  15. [23]

    Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. 2023. Spherical trans- former for lidar-based 3d recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 17545–17555

    2023

  16. [24]

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. 2022. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8500–8509

    2022

  17. [25]

    Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, and Federico Tombari. 2022. 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In Proceedin...

    2022

  18. [26]

    Huan Lei, Naveed Akhtar, and Ajmal Mian. 2020. Spherical kernel for efficient graph convolution on 3d point clouds. IEEE transactions on pattern analysis and machine intelligence 43, 10 (2020), 3664–3680

    2020

  19. [27]

    Guanlin Li, Guowen Xu, Han Qiu, Ruan He, Jiwei Li, and Tianwei Zhang. 2022. Improving adversarial robustness of 3D point cloud classification models. In European conference on computer vision . Springer, 672–689

    2022

  20. [28]

    Ruihui Li, Xianzhi Li, Pheng-Ann Heng, and Chi-Wing Fu. 2020. Pointaugment: an auto-augmentation framework for point cloud classification. (2020), 6378–6387

    2020

  21. [29]

    Sohee Lim, Minwoo Shin, and Joonki Paik. 2022. Point cloud generation using deep adversarial local features for augmented and mixed reality contents. IEEE Transactions on Consumer Electronics 68, 1 (2022), 69–76

    2022

  22. [30]

    Jianhua Lin. 1991. Divergence measures based on the Shannon entropy. IEEE Transactions on Information theory 37, 1 (1991), 145–151

    1991

  23. [31]

    Shuangjun Liu and Sarah Ostadabbas. 2018. A semi-supervised data augmentation approach using 3d graphical engines. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops . 0–0

    2018

  24. [32]

    Lingfei Ma, Ying Li, Jonathan Li, Weikai Tan, Yongtao Yu, and Michael A Chap- man. 2019. Multi-scale point-wise convolutional neural networks for 3D object segmentation from LiDAR point clouds in large-scale environments. IEEE Trans- actions on Intelligent Transportation Syste...

    2019

  25. [33]

    Alexey Nekrasov, Jonas Schult, Or Litany, Bastian Leibe, and Francis Engelmann

  26. [34]

    Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Hengshuang Zhao, Zhuotao Tian, and Jiaya Jia. 2024. Oa-cnns: Omni-adaptive sparse cnns for 3d semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21305–21315

    2024

  27. [35]

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. 2017. Pointnet: Deep learning on point sets for 3d classification and segmentation. (2017), 652–660

    2017

  28. [36]

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)

    2017

  29. [37]

    Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mo- hamed Elhoseiny, and Bernard Ghanem. 2022. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in neural information processing systems 35 (2022), 23192–23204

    2022

  30. [38]

    Yuan Ren, Siyan Zhao, and Liu Bingbing. 2022. Object insertion based data augmentation for semantic segmentation. In 2022 International Conference on Robotics and Automation (ICRA). IEEE, 359–365

    2022

  31. [39]

    Sushmita Sarker, Prithul Sarker, Gunner Stone, Ryan Gorman, Alireza Tavakkoli, George Bebis, and Javad Sattarvand. 2024. A comprehensive overview of deep learning techniques for 3D point cloud classification and semantic segmentation. Machine Vision and Applications 35, 4 (2024), 67

    2024

  32. [40]

    Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. 2020. Searching efficient 3d architectures with sparse point-voxel convolution. In European conference on computer vision . Springer, 685–702

    2020

  33. [41]

    Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. 2024. Diffuscene: Denoising diffusion models for generative indoor scene synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20507–20518

    2024

  34. [42]

    Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J Guibas. 2019. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international confer- ence on computer vision . 6411–6420

    2019

  35. [43]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  36. [44]

    Peng-Shuai Wang. 2023. Octformer: Octree-based transformers for 3d point clouds. ACM Transactions on Graphics (TOG) 42, 4 (2023), 1–11

    2023

  37. [45]

    Zaitian Wang, Pengfei Wang, Kunpeng Liu, Pengyang Wang, Yanjie Fu, Chang- Tien Lu, Charu C Aggarwal, Jian Pei, and Yuanchun Zhou. 2024. A comprehensive survey on data augmentation. arXiv preprint arXiv:2405.09591 (2024)

    2024

  38. [46]

    Liyana Wijayathunga, Alexander Rassau, and Douglas Chai. 2023. Challenges and solutions for autonomous ground robot scene understanding and navigation MM ’25, October 27–31, 2025, Dublin, Ireland H. Lin et al. in unstructured outdoor environments: A review. Applied Sciences 13...

    2023

  39. [47]

    Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. 2018. Squeezeseg: Con- volutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In 2018 IEEE international conference on robotics and automation (ICRA). IEEE, 1887–1893

    2018

  40. [48]

    Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. 2019. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In 2019 international conference on robotics and automation (ICRA) . IEEE...

    2019

  41. [49]

    Xiaoyang Wu, Daniel DeTone, Duncan Frost, Tianwei Shen, Chris Xie, Nan Yang, Jakob Engel, Richard Newcombe, Hengshuang Zhao, and Julian Straub

  42. [50]

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. 2024. Point transformer v3: Simpler faster stronger. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 4840–4851

    2024

  43. [51]

    Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. 2022. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems 35 (2022), 33330–33342

    2022

  44. [52]

    Zhengkang Xiang, Zexian Huang, and Kourosh Khoshelham. 2024. Synthetic lidar point cloud generation using deep generative models for improved driving scene object recognition. Image and Vision Computing 150 (2024), 105207

    2024

  45. [53]

    Aoran Xiao, Jiaxing Huang, Dayan Guan, Kaiwen Cui, Shijian Lu, and Ling Shao

  46. [54]

    Aoran Xiao, Xiaoqin Zhang, Ling Shao, and Shijian Lu. 2024. A survey of label- efficient deep learning for 3d point clouds. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    2024

  47. [55]

    Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. 2019. Pointflow: 3d point cloud generation with continuous normal- izing flows. (2019), 4541–4550

    2019

  48. [56]

    Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. 2025. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. Computational Visual Media 11, 1 (2025), 83–101

    2025

  49. [57]

    Chuangguan Ye, Hongyuan Zhu, Bo Zhang, and Tao Chen. 2023. A closer look at few-shot 3d point cloud classification. International Journal of Computer Vision 131, 3 (2023), 772–795

    2023

  50. [58]

    Guangyao Zhai, Evin Pınar Örnek, Dave Zhenyu Chen, Ruotong Liao, Yan Di, Nassir Navab, Federico Tombari, and Benjamin Busam. 2024. Echoscene: Indoor scene generation via information echo over scene graph diffusion. (2024), 167– 184

    2024

  51. [59]

    Guangyao Zhai, Evin Pınar Örnek, Shun-Cheng Wu, Yan Di, Federico Tombari, Nassir Navab, and Benjamin Busam. 2023. Commonscenes: Generating com- monsense 3d indoor scenes with scene graph diffusion. Advances in Neural Information Processing Systems 36 (2023), 30026–30038

    2023

  52. [60]

    Jinlai Zhang, Lyujie Chen, Bo Ouyang, Binbin Liu, Jihong Zhu, Yujin Chen, Yanmei Meng, and Danfeng Wu. 2022. Pointcutmix: Regularization strategy for point cloud classification. Neurocomputing 505 (2022), 58–67

    2022

  53. [61]

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. 2021. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision. 16259–16268

    2021

  54. [62]

    Chengyu Zheng, Mengjiao Ma, Zhilei Chen, Honghua Chen, Weiming Wang, and Mingqiang Wei. 2024. RegiFormer: Unsupervised Point Cloud Registra- tion via Geometric Local-to-Global Transformer and Self Augmentation. IEEE Transactions on Geoscience and Remote Sensing (2024)

    2024

  55. [63]

    Hui Zhou, Xinge Zhu, Xiao Song, Yuexin Ma, Zhe Wang, Hongsheng Li, and Dahua Lin. 2020. Cylinder3d: An effective 3d framework for driving-scene lidar semantic segmentation. arXiv preprint arXiv:2008.01550 (2020)

    2020 arXiv

  56. [64]

    Yang Zhou, Zachary While, and Evangelos Kalogerakis. 2019. Scenegraphnet: Neural message passing for 3d indoor scene augmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 7384–7392

    2019

  57. [65]

    WMSC testing mIoU

    Qinfeng Zhu, Lei Fan, and Ningxin Weng. 2024. Advancements in point cloud data augmentation for deep learning: A survey. Pattern Recognition (2024), 110532. Graph-Guided Dual-Level Augmentation for 3D Scene Segmentation MM ’25, October 27–31, 2025, Dublin, Ireland A Appendix A...

    2024

  58. [71]

    Node Sampling. Each object category’s activation probability is modeled using a Gaussian distribution, where the mean is set to the average number of instances of that category observed per scene in the training set. For example, if chairs appear 3 times on average in a scene,...

  59. [72]

    Once nodes are sampled, edges between all node pairs are activated based on the empirical relationship probability distribution obtained from the training set

    Edge Activation. Once nodes are sampled, edges between all node pairs are activated based on the empirical relationship probability distribution obtained from the training set. For instance, MM ’25, October 27–31, 2025, Dublin, Ireland H. Lin et al. Figure 6: Visualization of ...

    2025

  60. [73]

    This angular difference is calculated using the dot product formula: cos(𝛼) = ®𝑛𝑝·®𝑛𝐺𝑇 |®𝑛𝑝|·|®𝑛𝐺𝑇| (18) and we enforce the constraint𝛼 <𝜃

    The normal vector angle condition: The angular difference 𝛼 between the normal vector®𝑛𝑝 of point 𝑝 and the normal vector ®𝑛𝐺𝑇 of the GT boundary should less than𝜃. This angular difference is calculated using the dot product formula: cos(𝛼) = ®𝑛𝑝·®𝑛𝐺𝑇 |®𝑛𝑝|·|®𝑛𝐺𝑇| (18) and we ...

    2025

  61. [2019]

    SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In Proc. of the IEEE/CVF International Conf. on Computer Vision (ICCV)

  62. [2021]

    In 2021 interna- tional conference on 3d vision (3dv)

    Mix3d: Out-of-context data augmentation for 3d scenes. In 2021 interna- tional conference on 3d vision (3dv) . IEEE, 116–125

    2021

  63. [2022]

    Advances in Neural Information Processing Systems 35 (2022), 11035–11048

    Polarmix: A general data augmentation technique for lidar point clouds. Advances in Neural Information Processing Systems 35 (2022), 11035–11048

    2022

  64. [2024]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Graphdreamer: Compositional 3d scene synthesis from scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21295–21304

  65. [2025]

    arXiv preprint arXiv:2503.16429 (2025)

    Sonata: Self-Supervised Learning of Reliable Point Representations. arXiv preprint arXiv:2503.16429 (2025)

    2025 arXiv

Pith tools

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