Pith. sign in

REVIEW 3 major objections 5 minor 59 references

High-quality Pseudo-labeling for Point Cloud Segmentation with Scene-level Annotation

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 The paper proposes a weakly supervised point cloud segmentation framework that generates high-quality pseudo-labels from scene-level category tags by aligning 3D point features with 2D image features through contrastive distillation and…

desk verdict A solid, honestly reported method that deserves a serious referee, but the headline gains are confounded by Mix3D augmentation that the baselines didn't use. read the letter →

arxiv 2506.23227 v1 pith:FDYK2HU5 submitted 2025-06-29 cs.CV

Lunhao Duan , Shanshan Zhao , Xingxing Weng , Jing Zhang , Gui-Song Xia This is my paper
classification cs.CV
keywords scene-levelannotationweaklysupervisedpointcloudsegmentationpseudo-labelingcross-modalcontrastivedistillationregion-pointsemanticconsistencyindoor3Dteacher-studentlearning
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 tackles 3D point cloud semantic segmentation when the only supervision is a scene-level label: a list of object categories present in the scene, with no per-point annotations. The authors aim to show that such coarse labels can still yield accurate per-point pseudo-labels, which can then train a strong segmentation model. Their method combines two complementary signals: cross-modal contrastive alignment between 3D points and the 2D image pixels they project to, and a region-point semantic consistency constraint that averages teacher predictions inside unsupervised regions to guide student point predictions. On ScanNet v2 and S3DIS, the reported gains over prior scene-level methods are large (8–19 mIoU points), narrowing the gap to supervision that requires clicking individual points.

What carries the argument

The two load-bearing mechanisms are the cross-modal feature guidance module and the region-point semantic consistency (RPC) module. Cross-modal feature guidance uses the known camera parameters to establish point-pixel correspondences, then applies a point-wise contrastive loss in both directions (2D-to-3D and 3D-to-2D) so that corresponding point and pixel features become close while non-corresponding features are repelled. The RPC module operates in a teacher-student setup: an unsupervised shape extraction algorithm produces regions, teacher point predictions are averaged per region to form regional semantics, a per-category adaptive threshold (inspired by curriculum pseudo-labeling) filters low-confidence regional outputs, and the surviving regional labels supervise the student's point-level predictions. The teacher is updated by exponential moving average, and Mix3D augmentation is applied only to the student to enrich scene-level labels.

What would settle it

Measure the semantic purity of the unsupervised regions used by the RPC module on a test scene: if a large fraction of regions contain two or more ground-truth classes, then the region-averaged teacher output inherits the majority class and should mislead the point-level student. A direct experiment would compare pseudo-label mIoU when regions are ground-truth segments (the paper's Table 4 upper bound, 69.2) versus deliberately impure regions; if mIoU does not track region purity, the benefit attributed to region-point consistency is not coming from region quality.

Watch

Extended reading notes

Core claim

The central claim is that scene-level annotation alone can produce pseudo-labels whose quality approaches that of sparse point-level supervision, provided the pseudo-label generator uses both 2D-3D feature alignment and region-level semantic smoothing. Concretely, the paper shows that a point-wise contrastive distillation loss, which pulls each 3D point feature toward its corresponding pixel feature and pushes it away from non-corresponding pixels, substantially improves 3D feature discriminability under weak labels. On top of that, a region-point semantic consistency module partitions the point cloud into unsupervised regions, averages the teacher model's point-level predictions within each region, filters those regional predictions with a per-category adaptive threshold, and uses them to supervise the student model's point-level logits. The combination lifts pseudo-label mIoU from 25.1 to 48.2 on ScanNet training set and from 33.9 to 50.5 on S3DIS in the 3D branch, and the final second-stage models reach 46.9 mIoU on ScanNet validation, 46.8 on ScanNet test, and 47.4 on S3DIS Area 5 under scene-level annotation only.

Load-bearing premise

The region-point semantic consistency module assumes that unsupervised region partitions group points of the same semantic class, so averaging teacher predictions within a region yields reliable guidance rather than mixing different classes together.

Editorial extensions

If this is right

  • Scene-level annotation, which takes under a minute per scene, can replace sparse point-level labeling (about two minutes per scene) with only a modest drop in segmentation accuracy, reducing annotation cost substantially.
  • Cross-modal contrastive guidance from a pretrained 2D backbone transfers useful visual knowledge to the 3D network, so the choice of 2D backbone matters more for pseudo-label quality than scaling the 3D backbone.
  • The adaptive per-category threshold makes the RPC module robust to imperfect region partitioning, so the method does not hinge on any single oversegmentation algorithm.
  • Because the same two-stage pipeline is used for pseudo-label generation and final training, any further improvement in pseudo-label quality automatically improves the deployed segmentation model.
  • The appended experiments suggest the module can also refine pseudo-labels that come from projected 2D open-vocabulary segmentation models, not just scene tags.

Reading between the lines

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

  • Our inference: the remaining 21-point gap between using WyPR regions and ground-truth regions (48.2 vs 69.2 mIoU in Table 4) implies that improved unsupervised region purity would translate almost directly into better pseudo-labels, making region partitioning a high-leverage target for future work.
  • Our inference: the framework is better described as generic pseudo-label refinement than as a scene-level-specific method; the paper's own appendix shows it also improves noisy labels from OpenSeg and LSeg, so it could be applied to any weak or noisy per-point supervision source.
  • Our inference: applying the same contrastive alignment and region-voting consistency to outdoor LiDAR point clouds would require reworking the adaptive threshold to handle extreme class imbalance (e.g., drivable surface vs. bicycle), which the paper identifies as the main obstacle in its outdoor experiments.
  • Our inference: the 2D branch's role is not merely to provide features but to serve as a second, independently initialized view of the same scene; this suggests that a third modality or a differently initialized 3D teacher could play the same role, providing a testable extension.
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. This paper addresses weakly supervised 3D semantic segmentation when only scene-level category labels are available for each point cloud. The authors propose a two-stage pipeline: first, a teacher-student training scheme generates point-level pseudo-labels, using (i) a cross-modal feature guidance module that aligns 3D point features with corresponding 2D image pixels through contrastive losses in both directions, and (ii) a region-point semantic consistency module that averages teacher predictions inside unsupervised regions and uses an adaptive category-dependent threshold to supervise the student. These pseudo-labels are then used to train a segmentation network in a second stage. Experiments on ScanNet v2 and S3DIS report the best scene-level mIoU so far (46.9/46.8 on ScanNet and 47.4 on S3DIS test), with extensive ablation studies, per-class analyses, and additional experiments combining the approach with open-vocabulary 2D priors on indoor and outdoor datasets.

Significance. The contribution is potentially significant: scene-level annotation is much cheaper than point-level or sparse-point annotation, and the paper reports large margins over prior scene-level methods, narrowing the gap to weakly supervised methods that use several labeled points per class. The paper's strengths include release of code, systematic ablations of the loss components, thresholds, partitions, backbones, number of views, and an honest per-class breakdown, including appendices that quantify failure classes and test open-vocabulary initialization. However, the headline SOTA claim is not yet fully supported because the main comparison is confounded by Mix3D augmentation and because the final-metric ablations are missing; the paper also reports collapsed per-class performance on two ScanNet categories. These issues are fixable with additional controlled experiments, but they need to be addressed before the central claim can be accepted.

major comments (3)
  1. [Section 3.3, Table 2, Table 1] The main Table 1 comparison is not controlled for Mix3D augmentation. The Mix3D description in Section 3.3 states that mixed scenes are used for the student network, and Table 2 shows that including Mix3D improves pseudo-label mIoU from 48.2 to 50.1 on ScanNet and from 50.5 to 51.6 on S3DIS in the full module configuration. All Table 1 rows for the proposed method use this augmented first stage, whereas none of the compared methods (MPRM, WyPR, MIT, Xia et al.) use Mix3D, and no Table 1 row reports Ours without Mix3D. Because the second stage is trained on the first-stage pseudo-labels, the reported +8.8 mIoU over Xia et al. on ScanNet and +19.7 mIoU over MIT on S3DIS may be partially an augmentation effect. The 'Baseline' row in Table 1 also deserves clarification, since it outperforms MIT on S3DIS test (31.5 vs 27.7) despite using only classification losses. Please report final ScanNet validation and test and S3DIS Area 5 mIoU for Ours without Mix3D, and ideally rerun at least MIT/Xia et al. with the same Mix3D protocol, so that the gain can be attributed to the proposed modules.
  2. [Tables 1-2 and Section 4.4] The ablations in Section 4.4 and Table 2 measure pseudo-label mIoU on training sets, while the headline result in Table 1 is the final segmentation mIoU after second-stage training. No ablation table reports final validation/test mIoU for the key configurations (baseline, CMG, RPC, Mix3D, full). Consequently, it is not demonstrated that the pseudo-label improvements of individual components propagate to the final metric, and the contribution of each module to the reported SOTA numbers is not quantified. Please provide final mIoU for at least the main ablation configurations in Table 2.
  3. [Appendix C, Tables 13 and 14; Table 16] The method obtains 0.0 IoU on the 'picture' and 'counter' classes on the ScanNet validation and test sets, while the training-set pseudo-labels for these classes are 0.3 and 0.0 IoU (Table 16). This is a regression relative to MIT, which obtains 1.1 and 31.4 on validation for the same classes, and it directly undermines the paper's 'high-quality pseudo-labels' claim for these categories. The paper should either explain and mitigate this failure (for example, by analyzing the adaptive threshold's interaction with very rare classes) or explicitly state that the method sacrifices certain rare classes for overall mIoU.
minor comments (5)
  1. [Equation (4)] Equation (4) contains an unmatched parenthesis after `Y_hat_3D_t[p, c])`; the loss formula should be corrected.
  2. [Appendix B] The opening sentence of Appendix B is ungrammatical: 'Applying baseline method using, which uses only scene-level tags...' should be rewritten.
  3. [Throughout] There are several typos and inconsistencies: 'Pre-reained' in the Appendix A heading, 'T raining.' at the start of Section 4.2, and inconsistent spellings 'WYPR'/'WyPR' and 'Mink U-Net 18a'/'Mink U-Net 18A'.
  4. [Table 1] Table 1 reports no ScanNet test value for the Baseline row; please state whether the baseline was evaluated on the test server and, if not, why.
  5. [Table 15] Table 15 shows 0.0 IoU for 'beam' and 'board' for all methods including ours; a brief note would prevent readers from misreading this as a method-specific failure.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: pseudo-labels are generated from scene-level labels via external 2D guidance and consistency training, with no fitted parameter renamed as a prediction; the only overlapping-author reference is a non-load-bearing related-work pointer.

full rationale

The derivation chain is self-contained against external benchmarks. The core supervision comes from scene-level multi-label losses, point-pixel contrastive distillation (Eq. 1) using ImageNet-pretrained 2D features and known camera correspondences, and region-point consistency (Eqs. 2-8) that averages teacher point predictions within unsupervised regions and thresholds them with category-frequency-adaptive values. No parameter is fitted to validation or test mIoU: temperature tau=0.07, alpha=0.05, beta=1.0, and the base threshold eta are stated fixed choices, and the adaptive threshold in Eq. 6 depends only on the counts of categories in the generated pseudo-labels, not on target performance. The region guidance in Eq. 2 is a smoothed version of the teacher's own point predictions, so the method is a self-training/consistency regularizer rather than a derivation that equates output with input; Table 4's large gap between unsupervised partitions (48.2) and ground-truth partitions (69.2) shows the guidance is not forced by construction. The only overlapping-author citation is [7] Condaformer, cited in the introduction as an example of point cloud operations; it is not used to justify any load-bearing premise, uniqueness claim, or ansatz. The paper also discloses its empirical limitations, including 0.0 mIoU on 'picture' and 'counter' in Appendix C and sensitivity to partitioning in Table 4, which further indicates the pipeline is not tautological. The main-table comparison uses Mix3D augmentation not used by prior baselines; this is a benchmark-comparability concern, not circularity, and does not raise the circularity score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on domain assumptions about cross-modal correspondence, region partition quality, and the validity of the scene-level data augmentation. Several hyperparameters (alpha, beta, tau, eta, number of views) are chosen or tuned on validation data, but none are fitted to the test labels. No new entities are introduced. The most fragile assumption is that unsupervised regions align with semantic classes, as evidenced by the large gap to ground-truth partitions in Table 4.

free parameters (6)
  • loss weight alpha = 0.05
    Set in Sec 4.2 to weight the contrastive distillation losses.
  • loss weight beta = 1.0
    Set in Sec 4.2 to weight the region-point consistency losses.
  • temperature tau = 0.07
    Set in Sec 4.2 for the contrastive loss.
  • base threshold eta = not stated
    The adaptive threshold depends on a base eta; Table 8 presents results over eta values 0.0-0.9, but the final eta used in the main experiments is not specified in the paper.
  • number of 2D views = 6
    Ablation in Table 6 shows 6 views is optimal, so this value is selected based on validation results.
  • number of point-pixel pairs = 512
    Set in Sec 4.2 for sampling contrastive pairs.
assumptions (4)
  • domain assumption Accurate 2D-3D correspondences
    Sec 4.2 projects points to pixels using camera parameters; occlusions, depth errors, or calibration inaccuracies would create mismatched pairs for the contrastive loss.
  • domain assumption Unsupervised region partitions align with semantic boundaries
    Sec 3.3 uses an unsupervised shape extraction algorithm; Table 4 shows ground-truth partitions far outperform unsupervised ones (69.2 vs 48.2 mIoU), so the method relies on approximate semantic alignment of regions.
  • domain assumption Floor/wall/ceiling removal heuristic yields valid negative labels
    Sec 4.2 removes the bottom/top 0.5m of each scene and sets floor/ceiling/wall labels to 0; this assumes these structures appear only in those bands, which can be violated by stairs, lofts, or partitions.
  • domain assumption Teacher-student EMA improves stability
    Sec 3.3 uses an EMA teacher to produce regional guidance; this standard technique assumes teacher outputs are stable enough to supervise the student without introducing drift.

how reviews work

0 comments
Cite this review

Pith. "Pith review of High-quality Pseudo-labeling for Point Cloud Segmentation with Scene-level Annotation." pith.science (2026). https://pith.science/paper/FDYK2HU5

@misc{pith2026250623227,
  author       = {Pith},
  title        = {Pith review of: High-quality Pseudo-labeling for Point Cloud Segmentation with Scene-level Annotation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDYK2HU5}},
  note         = {Machine review of arXiv:2506.23227}
}
read the original abstract

This paper investigates indoor point cloud semantic segmentation under scene-level annotation, which is less explored compared to methods relying on sparse point-level labels. In the absence of precise point-level labels, current methods first generate point-level pseudo-labels, which are then used to train segmentation models. However, generating accurate pseudo-labels for each point solely based on scene-level annotations poses a considerable challenge, substantially affecting segmentation performance. Consequently, to enhance accuracy, this paper proposes a high-quality pseudo-label generation framework by exploring contemporary multi-modal information and region-point semantic consistency. Specifically, with a cross-modal feature guidance module, our method utilizes 2D-3D correspondences to align point cloud features with corresponding 2D image pixels, thereby assisting point cloud feature learning. To further alleviate the challenge presented by the scene-level annotation, we introduce a region-point semantic consistency module. It produces regional semantics through a region-voting strategy derived from point-level semantics, which are subsequently employed to guide the point-level semantic predictions. Leveraging the aforementioned modules, our method can rectify inaccurate point-level semantic predictions during training and obtain high-quality pseudo-labels. Significant improvements over previous works on ScanNet v2 and S3DIS datasets under scene-level annotation can demonstrate the effectiveness. Additionally, comprehensive ablation studies validate the contributions of our approach's individual components. The code is available at https://github.com/LHDuan/WSegPC .

Figures

Figures reproduced from arXiv: 2506.23227 by the authors.

Figure 1
Figure 1. Comparisons between pseudo-labels generated by PCAM [18] [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework: The point cloud and multi-view images are initially processed by the 3D and 2D branches to extract [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of semantic segmentation results on ScanNet v2 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of semantic segmentation results on the ScanNet and S3DIS training sets. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 52 canonical work pages

  1. [1]

    Kpconv: Flexible and deformable convolution for point clouds,

    H. Thomas, C. R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and L. J. Guibas, “Kpconv: Flexible and deformable convolution for point clouds,” in Int. Conf. Comput. Vis., 2019, pp. 6411–6420

    2019

  2. [2]

    Learning semantic segmentation of large-scale point clouds with random sampling,

    Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, and A. Markham, “Learning semantic segmentation of large-scale point clouds with random sampling,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 11, pp. 8338–8354, 2021

    work page 2021

  3. [3]

    Cylindrical and asymmetrical 3D convolution net- works for lidar-based perception,

    X. Zhu, H. Zhou, T. Wang, F. Hong, W. Li, Y. Ma, H. Li, R. Yang, and D. Lin, “Cylindrical and asymmetrical 3D convolution net- works for lidar-based perception,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 10, pp. 6807–6822, 2021

    work page 2021

  4. [4]

    Point trans- former,

    H. Zhao, L. Jiang, J. Jia, P . H. Torr, and V . Koltun, “Point trans- former,” in Int. Conf. Comput. Vis., 2021, pp. 16 259–16 268

    2021

  5. [5]

    Point transformer v2: Grouped vector attention and partition-based pooling,

    X. Wu, Y. Lao, L. Jiang, X. Liu, and H. Zhao, “Point transformer v2: Grouped vector attention and partition-based pooling,” inAdv. Neural Inform. Process. Syst., 2022

    work page 2022

  6. [6]

    Octformer: Octree-based transformers for 3D point clouds,

    P .-S. Wang, “Octformer: Octree-based transformers for 3D point clouds,” arXiv preprint arXiv:2305.03045, 2023

    arXiv 2023

  7. [7]

    Condaformer: Disassembled transformer with local structure en- hancement for 3D point cloud understanding,

    L. Duan, S. Zhao, N. Xue, M. Gong, G.-S. Xia, and D. Tao, “Condaformer: Disassembled transformer with local structure en- hancement for 3D point cloud understanding,” Adv. Neural Inform. Process. Syst., vol. 36, 2024

    work page 2024

  8. [8]

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

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3D classification and segmentation,” inIEEE Conf. Comput. Vis. Pattern Recog., 2017, pp. 652–660

    work page 2017

Show all 59 references
  1. [9]

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

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,” Adv. Neural Inform. Process. Syst., vol. 30, 2017

    2017

  2. [10]

    Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels,

    X. Xu and G. H. Lee, “Weakly supervised semantic point cloud segmentation: Towards 10x fewer labels,” in IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 13 706–13 715

    2020

  3. [11]

    Sqn: Weakly-supervised semantic segmentation of large-scale 3D point clouds,

    Q. Hu, B. Yang, G. Fang, Y. Guo, A. Leonardis, N. Trigoni, and A. Markham, “Sqn: Weakly-supervised semantic segmentation of large-scale 3D point clouds,” in Eur. Conf. Comput. Vis. , 2022, pp. 600–619

    2022

  4. [12]

    One thing one click: A self-training approach for weakly supervised 3D semantic segmentation,

    Z. Liu, X. Qi, and C.-W. Fu, “One thing one click: A self-training approach for weakly supervised 3D semantic segmentation,” in IEEE Conf. Comput. Vis. Pattern Recog., 2021, pp. 1726–1736

    2021

  5. [13]

    Perturbed self- distillation: Weakly supervised large-scale point cloud semantic segmentation,

    Y. Zhang, Y. Qu, Y. Xie, Z. Li, S. Zheng, and C. Li, “Perturbed self- distillation: Weakly supervised large-scale point cloud semantic segmentation,” in Int. Conf. Comput. Vis., 2021, pp. 15 520–15 528

    2021

  6. [14]

    An mil-derived transformer for weakly supervised point cloud segmentation,

    C.-K. Yang, J.-J. Wu, K.-S. Chen, Y.-Y. Chuang, and Y.-Y. Lin, “An mil-derived transformer for weakly supervised point cloud segmentation,” in IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 11 830–11 839

    2022

  7. [15]

    Cpcm: Contextual point cloud modeling for weakly- supervised point cloud semantic segmentation,

    L. Liu, Z. Zhuang, S. Huang, X. Xiao, T. Xiang, C. Chen, J. Wang, and M. Tan, “Cpcm: Contextual point cloud modeling for weakly- supervised point cloud semantic segmentation,” in Int. Conf. Com- put. Vis., 2023, pp. 18 413–18 422

    2023

  8. [16]

    All points matter: Entropy-regularized distribution alignment for weakly- supervised 3D segmentation,

    L. Tang, Z. Chen, S. Zhao, C. Wang, and D. Tao, “All points matter: Entropy-regularized distribution alignment for weakly- supervised 3D segmentation,” Adv. Neural Inform. Process. Syst. , vol. 36, 2024

    2024

  9. [17]

    Scannet: Richly-annotated 3D reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3D reconstructions of indoor scenes,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2017, pp. 5828–5839

    2017

  10. [18]

    Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds,

    J. Wei, G. Lin, K.-H. Yap, T.-Y. Hung, and L. Xie, “Multi-path region mining for weakly supervised 3D semantic segmentation on point clouds,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2020, pp. 4384– 4393

    2020

  11. [19]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” in IEEE Conf. Comput. Vis. Pattern Recog., 2016, pp. 2921–2929

    2016

  12. [20]

    3D spatial recognition without spatially labeled 3D,

    Z. Ren, I. Misra, A. G. Schwing, and R. Girdhar, “3D spatial recognition without spatially labeled 3D,” in IEEE Conf. Comput. Vis. Pattern Recog., 2021, pp. 13 204–13 213

    2021

  13. [21]

    Joint learning of 2D-3D weakly super- vised semantic segmentation,

    H. Kweon and K.-J. Yoon, “Joint learning of 2D-3D weakly super- vised semantic segmentation,” in Adv. Neural Inform. Process. Syst., 2022

    2022

  14. [22]

    2D-3D interlaced transformer for point cloud segmentation with scene- level supervision,

    C.-K. Yang, M.-H. Chen, Y.-Y. Chuang, and Y.-Y. Lin, “2D-3D interlaced transformer for point cloud segmentation with scene- level supervision,” in Int. Conf. Comput. Vis., 2023, pp. 977–987

    2023

  15. [23]

    Dimensionality reduction by learning an invariant mapping,

    R. Hadsell, S. Chopra, and Y. LeCun, “Dimensionality reduction by learning an invariant mapping,” in IEEE Conf. Comput. Vis. Pattern Recog., 2006, pp. 1735–1742

    2006

  16. [24]

    Representation learning with contrastive predictive coding,

    A. v. d. Oord, Y. Li, and O. Vinyals, “Representation learning with contrastive predictive coding,” arXiv preprint arXiv:1807.03748 , 2018

    2018 arXiv

  17. [25]

    Creating large-scale city models from 3D-point clouds: a robust approach with hybrid representation,

    F. Lafarge and C. Mallet, “Creating large-scale city models from 3D-point clouds: a robust approach with hybrid representation,” Int. J. Comput. Vis., 2012

    2012

  18. [26]

    3D semantic parsing of large-scale indoor spaces,

    I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fis- cher, and S. Savarese, “3D semantic parsing of large-scale indoor spaces,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2016, pp. 1534– 1543

    2016

  19. [27]

    Weakly supervised semantic segmentation for large-scale point cloud,

    Y. Zhang, Z. Li, Y. Xie, Y. Qu, C. Li, and T. Mei, “Weakly supervised semantic segmentation for large-scale point cloud,” in AAAI, vol. 35, no. 4, 2021, pp. 3421–3429

    2021

  20. [28]

    Hy- bridcr: Weakly-supervised 3D point cloud semantic segmentation via hybrid contrastive regularization,

    M. Li, Y. Xie, Y. Shen, B. Ke, R. Qiao, B. Ren, S. Lin, and L. Ma, “Hy- bridcr: Weakly-supervised 3D point cloud semantic segmentation via hybrid contrastive regularization,” in IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 14 930–14 939

    2022

  21. [29]

    Dual adaptive transfor- mations for weakly supervised point cloud segmentation,

    Z. Wu, Y. Wu, G. Lin, J. Cai, and C. Qian, “Dual adaptive transfor- mations for weakly supervised point cloud segmentation,” in Eur. Conf. Comput. Vis., 2022, pp. 78–96

    2022

  22. [30]

    Densify your labels: Unsupervised clustering with bipar- tite matching for weakly supervised point cloud segmentation,

    S. Xia, J. Yue, K. Kania, L. Fang, A. Tagliasacchi, K. M. Yi, and W. Sun, “Densify your labels: Unsupervised clustering with bipar- tite matching for weakly supervised point cloud segmentation,” arXiv preprint arXiv:2312.06799, 2023

    2023 arXiv

  23. [31]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  24. [32]

    Cross- modal learning for domain adaptation in 3D semantic segmen- tation,

    M. Jaritz, T.-H. Vu, R. De Charette, ´E. Wirbel, and P . P´erez, “Cross- modal learning for domain adaptation in 3D semantic segmen- tation,” IEEE Trans. Pattern Anal. Mach. Intell. , vol. 45, no. 2, pp. 1533–1544, 2022

    2022

  25. [33]

    Cross modal distillation for supervision transfer,

    S. Gupta, J. Hoffman, and J. Malik, “Cross modal distillation for supervision transfer,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2016, pp. 2827–2836

    2016

  26. [34]

    Learning from 2D: Con- trastive pixel-to-point knowledge transfer for 3D pretraining,

    Y.-C. Liu, Y.-K. Huang, H.-Y. Chiang, H.-T. Su, Z.-Y. Liu, C.-T. Chen, C.-Y. Tseng, and W. H. Hsu, “Learning from 2D: Con- trastive pixel-to-point knowledge transfer for 3D pretraining,” arXiv preprint arXiv:2104.04687, 2021

    2021 arXiv

  27. [35]

    Image-to-lidar self-supervised distillation for autonomous driv- ing data,

    C. Sautier, G. Puy, S. Gidaris, A. Boulch, A. Bursuc, and R. Marlet, “Image-to-lidar self-supervised distillation for autonomous driv- ing data,” in IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 9891– 9901

    2022

  28. [36]

    2Dpass: 2D priors assisted semantic segmentation on lidar point clouds,

    X. Yan, J. Gao, C. Zheng, C. Zheng, R. Zhang, S. Cui, and Z. Li, “2Dpass: 2D priors assisted semantic segmentation on lidar point clouds,” in Eur. Conf. Comput. Vis., 2022, pp. 677–695

    2022

  29. [37]

    Camliflow: bidirectional camera-lidar fusion for joint optical flow and scene flow estimation,

    H. Liu, T. Lu, Y. Xu, J. Liu, W. Li, and L. Chen, “Camliflow: bidirectional camera-lidar fusion for joint optical flow and scene flow estimation,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2022, pp. 5791–5801

    2022

  30. [38]

    Bridged transformer for vision and point cloud 3D object detec- tion,

    Y. Wang, T. Ye, L. Cao, W. Huang, F. Sun, F. He, and D. Tao, “Bridged transformer for vision and point cloud 3D object detec- tion,” in IEEE Conf. Comput. Vis. Pattern Recog. , 2022, pp. 12 114– 12 123

    2022

  31. [39]

    Bidirectional projection network for cross dimension scene understanding,

    W. Hu, H. Zhao, L. Jiang, J. Jia, and T.-T. Wong, “Bidirectional projection network for cross dimension scene understanding,” in IEEE Conf. Comput. Vis. Pattern Recog., 2021, pp. 14 373–14 382

    2021

  32. [40]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,

    B. Zhang, Y. Wang, W. Hou, H. WU, J. Wang, M. Okumura, and T. Shinozaki, “Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,” in Adv. Neural Inform. Process. Syst. , 2021

    2021

  33. [41]

    Mix3d: Out-of-context data augmentation for 3D scenes,

    A. Nekrasov, J. Schult, O. Litany, B. Leibe, and F. Engelmann, “Mix3d: Out-of-context data augmentation for 3D scenes,” inIEEE Int. Conf. 3D Vis., 2021, pp. 116–125

    2021

  34. [42]

    4d spatio-temporal convnets: Minkowski convolutional neural networks,

    C. Choy, J. Gwak, and S. Savarese, “4d spatio-temporal convnets: Minkowski convolutional neural networks,” in IEEE Conf. Comput. Vis. Pattern Recog., 2019, pp. 3075–3084

    2019

  35. [43]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in IEEE Conf. Comput. Vis. Pattern Recog., 2016, pp. 770–778

    2016

  36. [44]

    Im- agenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Im- agenet: A large-scale hierarchical image database,” in IEEE Conf. Comput. Vis. Pattern Recog., 2009, pp. 248–255

    2009

  37. [45]

    Voxel cloud connectivity segmentation-supervoxels for point clouds,

    J. Papon, A. Abramov, M. Schoeler, and F. Worgotter, “Voxel cloud connectivity segmentation-supervoxels for point clouds,” in IEEE Conf. Comput. Vis. Pattern Recog., 2013, pp. 2027–2034

    2013

  38. [46]

    Seeded region growing,

    R. Adams and L. Bischof, “Seeded region growing,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 16, no. 6, pp. 641–647, 1994

    1994

  39. [47]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y. Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 9729–9738

    2020

  40. [48]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P . Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” in Int. Conf. Comput. Vis., 2021, pp. 9650–9660

    2021

  41. [49]

    nuscenes: A mul- timodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. Vora, V . E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuscenes: A mul- timodal dataset for autonomous driving,” in IEEE Conf. Comput. Vis. Pattern Recog., 2020, pp. 11 621–11 631

    2020

  42. [50]

    Semantickitti: A dataset for semantic scene understanding of lidar sequences,

    J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stach- niss, and J. Gall, “Semantickitti: A dataset for semantic scene understanding of lidar sequences,” 2019, pp. 9297–9307

    2019

  43. [51]

    3D weakly supervised semantic segmentation with 2D vision-language guidance,

    X. Xu, Y. Yuan, J. Li, Q. Zhang, Z. Jie, L. Ma, H. Tang, N. Sebe, and X. Wang, “3D weakly supervised semantic segmentation with 2D vision-language guidance,” in Eur. Conf. Comput. Vis., 2024, pp. 87–104

    2024

  44. [52]

    Scaling open-vocabulary image segmentation with image-level labels,

    G. Ghiasi, X. Gu, Y. Cui, and T.-Y. Lin, “Scaling open-vocabulary image segmentation with image-level labels,” inEur. Conf. Comput. Vis., 2022, pp. 540–557

    2022

  45. [53]

    Language-driven semantic segmentation,

    B. Li, K. Q. Weinberger, S. Belongie, V . Koltun, and R. Ranftl, “Language-driven semantic segmentation,” in Int. Conf. Learn. Represent., 2022

    2022

  46. [54]

    Openscene: 3D scene understanding with open vocabularies,

    S. Peng, K. Genova, C. Jiang, A. Tagliasacchi, M. Pollefeys, T. Funkhouser et al. , “Openscene: 3D scene understanding with open vocabularies,” in IEEE Conf. Comput. Vis. Pattern Recog., 2023, pp. 815–824

    2023

  47. [55]

    Open vocabulary 3D scene understanding via geometry guided self- distillation,

    P . Wang, Y. Wang, S. Li, Z. Zhang, Z. Lei, and L. Zhang, “Open vocabulary 3D scene understanding via geometry guided self- distillation,” in Eur. Conf. Comput. Vis. Springer, 2024, pp. 442–460

    2024

  48. [56]

    Open-vocabulary 3D semantic segmentation with foundation models,

    L. Jiang, S. Shi, and B. Schiele, “Open-vocabulary 3D semantic segmentation with foundation models,” in IEEE Conf. Comput. Vis. Pattern Recog., 2024, pp. 21 284–21 294

    2024

  49. [57]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y. Lo et al. , “Segment anything,” in Int. Conf. Comput. Vis., 2023, pp. 4015–4026. APPENDIX A MORE ABLATION STUDIES Analysis of V arious 2D Image Views. Experiments were c...

    2023

  50. [58]

    For instance, ‘drivable surface’, ‘manmade’, and ‘veg- etation’ constitute over 70% of all points, while rare classes like ‘bicycle’ or ‘motorcycle’ represent less than 0.1%

    Extreme Class Imbalance: Outdoor scenes exhibit ex- treme imbalance in point distribution across categories. For instance, ‘drivable surface’, ‘manmade’, and ‘veg- etation’ constitute over 70% of all points, while rare classes like ‘bicycle’ or ‘motorcycle’ represent less than...

  51. [59]

    Baseline

    High Scene Label Co-occurrence: Many common out- door categories (e.g., ‘car’, ‘drivable surface’, ‘side- walk’, ‘manmade’, ‘vegetation’) co-occur in almost ev- ery scene (present in more than 95% of 28,130 scenes). This high degree of co-occurrence makes the scene- level labe...

Pith tools

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