Pith. sign in

REVIEW 4 major objections 5 minor 54 references

2D foundation model masks, back-projected into 3D scenes, turn 20 point labels per scene into dense training labels and lift ScanNetV2 mIoU from 63.9 to 69.4.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A weakly supervised 3D point cloud segmentation method that back-projects Semantic-SAM 2D masks into 3D, propagates sparse labels inside masks, and uses reliability-filtered pseudo labels, reporting state-of-the-art mIoU on ScanNetV2 and S3DIS.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid, incremental method paper: a new pipeline combining 2D foundation-model masks with pseudo-label expansion gives real gains over prior weak supervision baselines, but the low S3DIS label accuracy and the missing comparison to the closest prior work keep it from being fully convincing yet. the 4 major comments →

arxiv 2508.19909 v1 pith:R3CZSEJU submitted 2025-08-27 cs.CV

Integrating SAM Supervision for 3D Weakly Supervised Point Cloud Segmentation

Lechun You , Zhonghua Wu , Weide Liu , Xulei Yang , Jun Cheng , Wei Zhou , Bharadwaj Veeravalli , Guosheng Lin This is my paper
classification cs.CV
keywords weakly supervised 3D semantic segmentationpoint cloud segmentation2D-to-3D mask projectionSemantic-SAMpseudo-label expansionconsistency regularizationnoise-robust lossScanNetV2 and S3DIS
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that the missing ingredient in weakly supervised 3D point cloud segmentation is not more 3D annotations but better use of cheap 2D supervision. It proposes taking masks produced by the 2D foundation model Semantic-SAM, projecting them into each 3D scene using camera poses and depth, and merging overlapping masks into dense “3D masks.” Sparse human labels are then spread across these masks by majority vote, and reliable model predictions are spread the same way during training; the enlarged label set is treated as noisy and handled by a noise-robust loss. On ScanNetV2 with 20 labeled points per scene the method reaches 69.4 mIoU, versus 63.9 for the previous best weakly supervised method and 60.1 for the strong Point Transformer V3 baseline; on S3DIS Area 5 under one-click-per-instance supervision it reaches 64.4 versus 58.4. If correct, this shows that 2D foundation models can substantially close the gap to fully supervised 3D segmentation without extra 3D annotation effort.

Core claim

The paper's central claim is that a sparse set of 3D point labels can be amplified into dense training labels by importing 2D segmentation masks from Semantic-SAM, back-projecting them into each 3D scene through a camera-derived link matrix, and merging overlapping masks across views. The few human annotations are then spread to fill the 3D masks by assigning each mask the modal class of the labels inside it. During training, reliable pseudo labels selected by confidence and uncertainty are likewise spread over masks when their modal proportion exceeds a threshold η, and the resulting expanded labels are trained with a normalized noise-robust loss. On ScanNetV2 with 20 points per scene, the

What carries the argument

The central mechanism is a set of dense 3D masks obtained by projecting Semantic-SAM's 2D segmentation masks into 3D point clouds via a link matrix of camera intrinsics, extrinsics, and depth-consistency checks, then merging masks across views by overlap. These masks act as weak region proposals: sparse annotations are extended by assigning each mask the mode of its annotated points, and later reliable pseudo labels are extended to an entire mask when the modal pseudo-label fraction exceeds a threshold η. The enlarged label set is treated as noisy and supervised with normalized cross-entropy plus reverse cross-entropy losses, which lets the pipeline absorb projection and boundary errors.

Load-bearing premise

The whole pipeline depends on the back-projected 2D masks matching true 3D object regions well enough that assigning each mask a single class label yields accurate training labels; if the masks cut objects apart, merge different objects, or misalign after projection, the expanded labels become systematically noisy and the reported gains would not transfer.

What would settle it

A reader could take the ScanNetV2 validation set with full ground truth, run the mask projection and label-propagation pipeline from a 20-point-per-scene annotation budget, and compare the expanded labels' per-mask accuracy against the claimed 93.0%. If masks frequently span multiple classes or projection misalignment causes modal labels to be wrong, the mIoU gain should disappear; that would show the 2D-mask expansion, rather than the consistency regularization, is not doing the claimed work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Annotation effort for 3D semantic segmentation can be cut to tens of points per scene (or one click per instance) while staying within 4–9 mIoU of fully supervised training on the same backbone.
  • The 2D masks are needed only at training time; inference remains pure 3D point cloud, so the method applies to scenes for which no RGB images are available at test time.
  • The label-expansion machinery is backbone-agnostic: the paper demonstrates gains on both Point Transformer V3 and SparseUNet.
  • Treating expanded labels as noisy and using normalized cross-entropy/reverse cross-entropy is what lets the pipeline absorb mask-boundary errors; removing that loss costs several mIoU in the ablations.
  • Thresholding reliable-pseudo-label expansion by their proportion in a mask (η) controls the trade-off between label quantity and label accuracy, so the optimal setting is dataset-dependent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A likely untested consequence is that replacing Semantic-SAM with other 2D segmenters should change final mIoU roughly in proportion to the quality of the back-projected masks, making this pipeline a direct benchmark for how much 2D mask quality matters for 3D label efficiency.
  • For outdoor autonomous-driving data, where RGB and LiDAR are captured together, the same projection-and-propagation pipeline could amplify sparse labels, but depth-based occlusion handling would have to cope with larger scale and moving objects.
  • The reported 93.0% accuracy of expanded ScanNet labels suggests remaining error concentrates at mask boundaries and small objects; trimming mask borders before propagation might further improve mIoU without new annotations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a weakly supervised 3D point cloud semantic segmentation method that leverages 2D foundation model masks. Using Semantic-SAM, it generates 2D segmentation masks for RGB views, back-projects them into 3D via a link matrix, merges overlapping 3D masks across views, and propagates both the initial sparse 3D annotations and reliable pseudo-labels onto the resulting masks. The model is trained with a combination of cross-entropy, consistency regularization (reliable/ambiguous split), KL divergence, and a noise-robust normalized loss. Experiments on ScanNetV2 (20 points per scene) and S3DIS (OTOC setting) with PTv3 and SparseUNet backbones report mIoU improvements over RAC-Net and the PTv3/SparseUNet baselines, with ablations showing each loss component contributes. The central claim is that back-projected 2D masks substantially expand the label pool and improve weakly supervised segmentation performance.

Significance. If the reported results hold, the paper offers a practical and timely method for tapping into 2D foundation models to reduce 3D annotation cost. The systematic ablations, quantitative characterization of expanded-label accuracy, and comparison with SAM3D-derived masks are strengths; the paper is reproducible in structure and reports concrete numbers. However, the strength of the central claim is constrained by the omission of the most closely related prior work [27], the absence of repeated-run statistics, and only a partial characterization of the noise introduced by mask-based label expansion, particularly on S3DIS where expanded-label accuracy is 71.6%. These are addressable but currently limit confidence in the claimed state-of-the-art margin.

major comments (4)
  1. [Section II-B, Tables I-II] The closest prior method using SAM masks for weakly supervised 3D segmentation, Dong et al. [27], is described in the related work but is not included in any experimental comparison. Since [27] also propagates sparse labels through 2D masks, the claim of 'state-of-the-art' among methods that exploit 2D foundation models cannot be assessed without benchmarking it under the same 20-point/OTOC protocols. Please add results for [27] or explicitly justify its exclusion.
  2. [Tables I-III] All reported results appear to come from a single run. With only 20 randomly sampled annotated points per scene on ScanNetV2 and one random OTOC point per instance on S3DIS, the reported margins over baselines (5.5–9.7 mIoU) could be within run-to-run variance. Please report mean and standard deviation over at least three independent annotation subsets/seeds for the main comparisons and key ablations.
  3. [Table IV, Section III-C] Expanded labels for S3DIS have only 71.6% accuracy, implying roughly 28% of the added training labels are wrong. The S3DIS gain of 9.7 mIoU over baseline is attributed to this expansion, yet the paper does not quantify how this noise level affects final performance. Please include an oracle experiment in which masks are labeled using ground truth, and a per-class accuracy analysis of the expanded labels eY, to demonstrate that the gain is not driven by class imbalance or systematic noise.
  4. [Section IV-C3, Table VIII, Algorithm 1] The key hyperparameter η is tuned on the ScanNetV2 validation set (Table VIII) and then apparently reused for S3DIS without sensitivity analysis. Other important hyperparameters—depth threshold δ, mask-overlap threshold, τ, κ, Nview, and loss weights—are not analyzed. Please report η sensitivity on both datasets and state the selection procedure for all hyperparameters, since the method's performance could hinge on matching them to the validation protocol.
minor comments (5)
  1. [Section III-A, Eq. (5)] The dimension M is described as 'the number of classes', but Semantic-SAM is a class-agnostic mask generator. Please clarify whether M denotes the number of semantic classes or the number of masks, and explain how class-agnostic masks are aligned with the 3D label set.
  2. [Section III-D, Eq. (11)] The NCE expression is hard to parse as typeset; the denominator appears to have a form that may be missing indices. Please verify that it matches the normalized loss of Ma et al. [24] and define q(k|x) clearly.
  3. [Algorithm 1] The pseudocode does not initialize eY and does not specify the output for a mask that has neither reliable pseudo-labels satisfying η nor any initial annotation. Please add an initialization and a default fallback.
  4. [Table IV caption] The caption says the average number of baseline annotations is 15 even though the setting is '20 points per scene'; please clarify why some randomly retained points lack ground truth.
  5. [Figure 1] The pipeline diagram is dense; consider labeling the four loss terms and the data flow for eY more explicitly so that the relationship to Eq. (13) is clearer.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical mIoU comparison on held-out benchmarks, not a derivation forced by definitions or self-citation.

full rationale

The paper's claimed contribution is a weakly supervised 3D segmentation pipeline whose main evidence is measured mIoU on held-out ScanNetV2 test and S3DIS Area5, compared against baselines and previous methods. The label expansion procedure (Algorithm 1) propagates initial labels and pseudo labels through back-projected 2D masks; this is a training-label construction, and the reported improvement is not an identity or a fitted parameter renamed as a prediction. Hyperparameters such as η and the choice of NCE/RCE loss are selected on the validation set and then applied, which is standard model selection rather than circular. The consistency-regularization component is taken from the authors' own RAC-Net [18] and is cited as methodology, but it is not used to establish the validity of the result; the result is evaluated against external benchmarks and the method is also tested on the SparseUNet backbone. The paper explicitly acknowledges the noise in expanded labels (e.g., 71.6% label accuracy on S3DIS, Table IV) and treats it as a problem to be mitigated with a noise-robust loss, which is an honest limitation rather than a circular step. No self-citation chain is load-bearing, no uniqueness theorem is imported, and no ansatz is smuggled in via citation to ground the conclusion.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical or data entities are invented; the method introduces algorithmic components (3D masks, reliable pseudo labels) rather than entities. The ledger lists the manually chosen hyperparameters and domain assumptions that the empirical result depends on.

free parameters (7)
  • eta (pseudo-label expansion ratio) = 0.7
    Chosen by validation sweep on ScanNetV2 (Table VIII); controls whether reliable pseudo labels are spread to an entire 3D mask.
  • tau (confidence threshold)
    Threshold in Eq. (7) for classifying predictions as reliable; value not reported.
  • kappa (uncertainty threshold)
    Threshold in Eq. (7) for classifying predictions as reliable; value not reported.
  • delta (depth matching threshold)
    Threshold in Eq. (4) for occlusion-aware 2D-to-3D projection; value not reported.
  • mask overlap threshold
    Threshold for merging projected masks from different views; value not reported.
  • Nview (number of sampled views)
    Number of 2D views uniformly sampled per scene; value not reported.
  • loss weights lambda_seg, lambda_r, lambda_a, lambda_m
    Weights in Eq. (13) balancing segmentation, reliable, ambiguous, and mask losses; values not reported.
axioms (5)
  • domain assumption Back-projection through camera matrix and depth resolves 2D-to-3D correspondences accurately enough for mask propagation.
    Section III-A, Eq. (1)-(4). The link matrix L depends on camera intrinsics/extrinsics and depth matching; any calibration error shifts masks to the wrong 3D points.
  • domain assumption Semantic-SAM masks partition pixels into object-coherent regions, so a 3D mask can be assigned a single class label.
    Section III-A and III-C. Label propagation via mode labels inside masks only makes sense if masks are class-pure.
  • ad hoc to paper Masks from different views that overlap by more than a threshold belong to the same class and should be merged.
    Section III-A merging paragraph. The threshold is not reported, and merging different objects would directly corrupt eY.
  • domain assumption Confidence and variance of predictions across augmentations identify accurate pseudo labels.
    Section III-B Eq. (7) follows RAC-Net [18]; reliable pseudo labels are subsequently expanded to full masks in Algorithm 1.
  • domain assumption NCE + RCE normalized loss is robust enough to train on the noisy expanded labels eY.
    Section III-D Eq. (10)-(11), borrowed from [24] and selected empirically in Table VI.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Integrating SAM Supervision for 3D Weakly Supervised Point Cloud Segmentation." pith.science (2026). https://pith.science/paper/R3CZSEJU

@misc{pith2026250819909,
  author       = {Pith},
  title        = {Pith review of: Integrating SAM Supervision for 3D Weakly Supervised Point Cloud Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R3CZSEJU}},
  note         = {Machine review of arXiv:2508.19909}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Current methods for 3D semantic segmentation propose training models with limited annotations to address the difficulty of annotating large, irregular, and unordered 3D point cloud data. They usually focus on the 3D domain only, without leveraging the complementary nature of 2D and 3D data. Besides, some methods extend original labels or generate pseudo labels to guide the training, but they often fail to fully use these labels or address the noise within them. Meanwhile, the emergence of comprehensive and adaptable foundation models has offered effective solutions for segmenting 2D data. Leveraging this advancement, we present a novel approach that maximizes the utility of sparsely available 3D annotations by incorporating segmentation masks generated by 2D foundation models. We further propagate the 2D segmentation masks into the 3D space by establishing geometric correspondences between 3D scenes and 2D views. We extend the highly sparse annotations to encompass the areas delineated by 3D masks, thereby substantially augmenting the pool of available labels. Furthermore, we apply confidence- and uncertainty-based consistency regularization on augmentations of the 3D point cloud and select the reliable pseudo labels, which are further spread on the 3D masks to generate more labels. This innovative strategy bridges the gap between limited 3D annotations and the powerful capabilities of 2D foundation models, ultimately improving the performance of 3D weakly supervised segmentation.

Figures

Figures reproduced from arXiv: 2508.19909 by Bharadwaj Veeravalli, Guosheng Lin, Jun Cheng, Lechun You, Weide Liu, Wei Zhou, Xulei Yang, Zhonghua Wu.

Figure 1
Figure 1. Figure 1: Our method consists of four main components. The Mask Generation and Spatial Mask Projection modules are to generate segmentation masks for [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of 3D labels from ScanNetV2 dataset before and after label initialization, with 20 labeled points per scene. The 3D masks projected [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of 3D labels from S3DIS dataset before and after label initialization with ‘OTOC’ annotation scheme. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A failure case from the ScanNetV2 dataset after label initialization. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Number and accuracy of Ye with different value of η through the training process. Corresponding to these computational entities, [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: An example of expanded labels from the ScanNetV2 dataset. The [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples of semantic segmentation results on ScanNetV2 validation [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of semantic segmentation results on S3DIS testing set (Area [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 36 canonical work pages · 13 internal anchors

  1. [1]

    PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,

    R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, Jul. 2017, pp. 77–85. [Online]. Available: http://ieeexplore.ieee.org/document/8099499/

    work page arXiv 2017

  2. [2]

    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,” Jun. 2017, arXiv:1706.02413 [cs]. [Online]. Available: http://arxiv.org/abs/1706.02413

    Pith/arXiv arXiv 2017

  3. [3]

    PointCNN: Convolution On X-Transformed Points,

    Y . Li, R. Bu, M. Sun, W. Wu, X. Di, and B. Chen, “PointCNN: Convolution On X-Transformed Points,” in Advances in Neural Information Processing Systems , vol. 31. Curran Associates, Inc.,

    work page

  4. [4]

    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,” Aug. 2019, arXiv:1904.08889 [cs]. [Online]. Available: http://arxiv.org/abs/1904.08889

    Pith/arXiv arXiv 2019

  5. [5]

    PointConv: Deep Convolutional Networks on 3D Point Clouds,

    W. Wu, Z. Qi, and L. Fuxin, “PointConv: Deep Convolutional Networks on 3D Point Clouds,” Nov. 2020, arXiv:1811.07246 [cs]. [Online]. Available: http://arxiv.org/abs/1811.07246

    Pith/arXiv arXiv 2020

  6. [6]

    Point Transformer,

    H. Zhao, L. Jiang, J. Jia, P. Torr, and V . Koltun, “Point Transformer,” Sep. 2021, arXiv:2012.09164 [cs]. [Online]. Available: http://arxiv.org/abs/2012.09164

    Pith/arXiv arXiv 2021

  7. [7]

    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,” Oct. 2022, arXiv:2210.05666 [cs]. [Online]. Available: http://arxiv.org/abs/2210. 05666

    Pith/arXiv arXiv 2022

  8. [8]

    3D Semantic Segmentation with Submanifold Sparse Convolutional Networks,

    B. Graham, M. Engelcke, and L. V . D. Maaten, “3D Semantic Segmentation with Submanifold Sparse Convolutional Networks,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA: IEEE, Jun. 2018, pp. 9224–9232. [Online]. Available: https://ieeexplore.ieee.org/document/ 8579059/

    work page 2018

  9. [9]

    4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks,

    C. Choy, J. Gwak, and S. Savarese, “4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural Networks,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Long Beach, CA, USA: IEEE, Jun. 2019, pp. 3070–3079. [Online]. Available: https://ieeexplore.ieee.org/document/8953494/ JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGU...

    work page arXiv 2019

  10. [10]

    SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud,

    B. Wu, A. Wan, X. Yue, and K. Keutzer, “SqueezeSeg: Convolutional 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) . Brisbane, QLD: IEEE, May 2018, pp. 1887–1893. [Online]. Available: https://ieeexplore.ieee.org/document/8462926/

    work page arXiv 2018

  11. [11]

    3DFeat-Net: Weakly Supervised Local 3D Features for Point Cloud Registration,

    Z. J. Yew and G. H. Lee, “3DFeat-Net: Weakly Supervised Local 3D Features for Point Cloud Registration,” in Computer Vision – ECCV 2018 , V . Ferrari, M. Hebert, C. Sminchisescu, and Y . Weiss, Eds. Cham: Springer International Publishing, 2018, vol. 11219, pp. 630–646, series Title: Lecture Notes in Computer Science. [Online]. Available: https://link.spr...

    doi:10.1007/978-3-030-01267-0 2018

  12. [12]

    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 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . Seattle, W A, USA: IEEE, Jun. 2020, pp. 4383–4392. [Online]. Available: https://ieeexplore.ieee.org/document/ 9157503/

    work page 2020

  13. [13]

    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 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN, USA: IEEE, Jun. 2021, pp. 1726–1736. [Online]. Available: https://ieeexplore.ieee.org/document/9578763/

    work page arXiv 2021

  14. [14]

    Dual Adaptive Transformations for Weakly Supervised Point Cloud Segmentation

    Z. Wu, Y . Wu, G. Lin, J. Cai, and C. Qian, “Dual Adaptive Transformations for Weakly Supervised Point Cloud Segmentation,” Jul. 2022, arXiv:2207.09084 [cs]. [Online]. Available: http://arxiv.org/ abs/2207.09084

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  15. [15]

    SegGroup: Seg-Level Supervision for 3D Instance and Semantic Segmentation

    A. Tao, Y . Duan, Y . Wei, J. Lu, and J. Zhou, “SegGroup: Seg-Level Supervision for 3D Instance and Semantic Segmentation,” IEEE Transactions on Image Processing , vol. 31, pp. 4952–4965, 2022, arXiv:2012.10217 [cs]. [Online]. Available: http://arxiv.org/abs/2012. 10217

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  16. [16]

    You Only Need One Thing One Click: Self-Training for Weakly Supervised 3D Scene Understanding

    Z. Liu, X. Qi, and C.-W. Fu, “One Thing One Click++: Self- Training for Weakly Supervised 3D Scene Understanding,” Mar. 2023, arXiv:2303.14727 [cs]. [Online]. Available: http://arxiv.org/abs/2303. 14727

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    PointMatch: A consistency training framework for weakly supervised semantic segmentation of 3D point clouds,

    Y . Wu, Z. Yan, S. Cai, G. Li, X. Han, and S. Cui, “PointMatch: A consistency training framework for weakly supervised semantic segmentation of 3D point clouds,” Computers & Graphics , vol. 116, pp. 427–436, Nov. 2023. [Online]. Available: https://linkinghub.elsevier. com/retrieve/pii/S0097849323002297

    work page 2023

  18. [18]

    Reliability-Adaptive Consistency Regularization for Weakly-Supervised Point Cloud Segmentation,

    Z. Wu, Y . Wu, G. Lin, and J. Cai, “Reliability-Adaptive Consistency Regularization for Weakly-Supervised Point Cloud Segmentation,” International Journal of Computer Vision , Jan. 2024. [Online]. Available: https://link.springer.com/10.1007/s11263-023-01975-8

    work page doi:10.1007/s11263-023-01975-8 2024

  19. [19]

    Virtual Multi-view Fusion for 3D Semantic Segmentation

    A. Kundu, X. Yin, A. Fathi, D. Ross, B. Brewington, T. Funkhouser, and C. Pantofaru, “Virtual Multi-view Fusion for 3D Semantic Segmentation,” Jul. 2020, arXiv:2007.13138 [cs, eess]. [Online]. Available: http://arxiv.org/abs/2007.13138

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  20. [20]

    3D-MiniNet: Learning a 2D Representation From Point Clouds for Fast and Efficient 3D LIDAR Semantic Segmentation,

    I. Alonso, L. Riazuelo, L. Montesano, and A. C. Murillo, “3D-MiniNet: Learning a 2D Representation From Point Clouds for Fast and Efficient 3D LIDAR Semantic Segmentation,” IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 5432–5439, Oct. 2020. [Online]. Available: https://ieeexplore.ieee.org/document/9134888/

    work page arXiv 2020

  21. [21]

    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,” Mar. 2021, arXiv:2103.14326 [cs]. [Online]. Available: http://arxiv.org/abs/ 2103.14326

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  22. [22]

    Learning 3D Semantic Segmentation with only 2D Image Supervision,

    K. Genova, X. Yin, A. Kundu, C. Pantofaru, F. Cole, A. Sud, B. Brewington, B. Shucker, and T. Funkhouser, “Learning 3D Semantic Segmentation with only 2D Image Supervision,” in 2021 International Conference on 3D Vision (3DV) . London, United Kingdom: IEEE, Dec. 2021, pp. 361–372. [Online]. Available: https://ieeexplore.ieee.org/document/9665849/

    work page arXiv 2021

  23. [23]

    Semantic-SAM: Segment and Recognize Anything at Any Granularity,

    F. Li, H. Zhang, P. Sun, X. Zou, S. Liu, J. Yang, C. Li, L. Zhang, and J. Gao, “Semantic-SAM: Segment and Recognize Anything at Any Granularity,” Jul. 2023, arXiv:2307.04767 [cs]. [Online]. Available: http://arxiv.org/abs/2307.04767

    Pith/arXiv arXiv 2023

  24. [24]

    Normalized Loss Functions for Deep Learning with Noisy Labels

    X. Ma, H. Huang, Y . Wang, S. Romano, S. Erfani, and J. Bailey, “Normalized Loss Functions for Deep Learning with Noisy Labels,” Jun. 2020, arXiv:2006.13554 [cs, stat]. [Online]. Available: http: //arxiv.org/abs/2006.13554

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  25. [25]

    Point Transformer V3: Simpler, Faster, Stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point Transformer V3: Simpler, Faster, Stronger,” Mar. 2024, arXiv:2312.10035 [cs]. [Online]. Available: http://arxiv.org/abs/2312.10035

    Pith/arXiv arXiv 2024

  26. [26]

    Active self-training for weakly supervised 3D scene semantic segmentation,

    G. Liu, O. Van Kaick, H. Huang, and R. Hu, “Active self-training for weakly supervised 3D scene semantic segmentation,” Computational Visual Media , vol. 10, no. 3, pp. 425–438, Jun. 2024. [Online]. Available: https://link.springer.com/10.1007/s41095-022-0311-7

    work page doi:10.1007/s41095-022-0311-7 2024

  27. [27]

    Leveraging Large-Scale Pretrained Vision Foundation Models for Label-Efficient 3D Point Cloud Segmentation,

    S. Dong, F. Liu, and G. Lin, “Leveraging Large-Scale Pretrained Vision Foundation Models for Label-Efficient 3D Point Cloud Segmentation,” Nov. 2023, arXiv:2311.01989 [cs]. [Online]. Available: http://arxiv.org/abs/2311.01989

    Pith/arXiv arXiv 2023

  28. [28]

    Segment Anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll ´ar, and R. Girshick, “Segment Anything,” Apr. 2023, arXiv:2304.02643 [cs]. [Online]. Available: http://arxiv.org/abs/2304.02643

    Pith/arXiv arXiv 2023

  29. [29]

    Crnet: Cross-reference networks for few-shot segmentation,

    W. Liu, C. Zhang, G. Lin, and F. Liu, “Crnet: Cross-reference networks for few-shot segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 4165–4173

    work page 2020

  30. [30]

    Few-shot segmentation with optimal transport matching and message flow,

    W. Liu, C. Zhang, H. Ding, T.-Y . Hung, and G. Lin, “Few-shot segmentation with optimal transport matching and message flow,” IEEE Transactions on Multimedia , vol. 25, pp. 5130–5141, 2022

    work page 2022

  31. [31]

    Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation,

    W. Liu, Z. Wu, Y . Zhao, Y . Fang, C.-S. Foo, J. Cheng, and G. Lin, “Harmonizing base and novel classes: A class-contrastive approach for generalized few-shot segmentation,” International Journal of Computer Vision, vol. 132, no. 4, pp. 1277–1291, 2024

    work page 2024

  32. [32]

    Modality-Aware Feature Matching in Visual and Vision-Language Applications: A Comprehensive Survey

    W. Liu, W. Zhou, J. Liu, P. Hu, J. Cheng, J. Han, and W. Lin, “Modality- aware feature matching: A comprehensive review of single-and cross- modality techniques,” arXiv preprint arXiv:2507.22791 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . Boston, MA, USA: IEEE, Jun. 2015, pp. 3431–3440. [Online]. Available: http://ieeexplore.ieee.org/document/7298965/

    work page arXiv 2015

  34. [34]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” May 2015, arXiv:1505.04597 [cs]. [Online]. Available: http://arxiv.org/abs/1505. 04597

    Pith/arXiv arXiv 2015

  35. [35]

    Conditional Random Fields as Recurrent Neural Networks,

    S. Zheng, S. Jayasumana, B. Romera-Paredes, V . Vineet, Z. Su, D. Du, C. Huang, and P. H. S. Torr, “Conditional Random Fields as Recurrent Neural Networks,” in 2015 IEEE International Conference on Computer Vision (ICCV) . Santiago, Chile: IEEE, Dec. 2015, pp. 1529–1537. [Online]. Available: http://ieeexplore.ieee.org/document/7410536/

    work page arXiv 2015

  36. [36]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” Jun. 2021, arXiv:2010.11929 [cs]. [Online]. Available: http://arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2021

  37. [37]

    SegViT: Semantic Segmentation with Plain Vision Transformers

    B. Zhang, Z. Tian, Q. Tang, X. Chu, X. Wei, C. Shen, and Y . Liu, “SegViT: Semantic Segmentation with Plain Vision Transformers,” Dec. 2022, arXiv:2210.05844 [cs]. [Online]. Available: http://arxiv.org/ abs/2210.05844

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  38. [38]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE, Oct. 2021, pp. 9992–10 002. [Online]. Available: https://ieeexplore.ieee.org/document/ 9710580/

    work page 2021

  39. [39]

    SAM3D: Segment Anything in 3D Scenes,

    Y . Yang, X. Wu, T. He, H. Zhao, and X. Liu, “SAM3D: Segment Anything in 3D Scenes,” Jun. 2023, arXiv:2306.03908. [Online]. Available: http://arxiv.org/abs/2306.03908

    Pith/arXiv arXiv 2023

  40. [40]

    Hierarchical Open-vocabulary Universal Image Segmentation

    X. Wang, S. Li, K. Kallidromitis, Y . Kato, K. Kozuka, and T. Darrell, “Hierarchical Open-vocabulary Universal Image Segmentation,” Jul. 2023, arXiv:2307.00764 [cs]. [Online]. Available: http://arxiv.org/abs/ 2307.00764

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  41. [41]

    Robust Loss Functions under Label Noise for Deep Neural Networks,

    A. Ghosh, H. Kumar, and P. S. Sastry, “Robust Loss Functions under Label Noise for Deep Neural Networks,” Dec. 2017, arXiv:1712.09482 [cs, stat]. [Online]. Available: http://arxiv.org/abs/1712.09482

    Pith/arXiv arXiv 2017

  42. [42]

    Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels,

    Z. Zhang and M. R. Sabuncu, “Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels,” Nov. 2018, arXiv:1805.07836 [cs, stat]. [Online]. Available: http://arxiv.org/abs/ 1805.07836

    Pith/arXiv arXiv 2018

  43. [43]

    Symmetric Cross Entropy for Robust Learning with Noisy Labels,

    Y . Wang, X. Ma, Z. Chen, Y . Luo, J. Yi, and J. Bailey, “Symmetric Cross Entropy for Robust Learning with Noisy Labels,” Aug. 2019, arXiv:1908.06112 [cs, stat]. [Online]. Available: http: //arxiv.org/abs/1908.06112

    Pith/arXiv arXiv 2019

  44. [44]

    Asymmetric Loss Functions for Learning with Noisy Labels

    X. Zhou, X. Liu, J. Jiang, X. Gao, and X. Ji, “Asymmetric Loss Functions for Learning with Noisy Labels,” Jun. 2021, arXiv:2106.03110 [cs]. [Online]. Available: http://arxiv.org/abs/2106.03110 JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  45. [45]

    Unsupervised Label Noise Modeling and Loss Correction,

    E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuinness, “Unsupervised Label Noise Modeling and Loss Correction,” Jun. 2019, arXiv:1904.11238 [cs]. [Online]. Available: http://arxiv.org/abs/1904. 11238

    Pith/arXiv arXiv 2019

  46. [46]

    Learning with Noisy Labels for Robust Point Cloud Segmentation,

    S. Ye, D. Chen, S. Han, and J. Liao, “Learning with Noisy Labels for Robust Point Cloud Segmentation,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) . Montreal, QC, Canada: IEEE, Oct. 2021, pp. 6423–6432. [Online]. Available: https://ieeexplore.ieee.org/document/9710169/

    work page arXiv 2021

  47. [47]

    Point Cloud Augmentation with Weighted Local Transformations,

    S. Kim, S. Lee, D. Hwang, J. Lee, S. J. Hwang, and H. J. Kim, “Point Cloud Augmentation with Weighted Local Transformations,” in 2021 IEEE/CVF International Conference on Computer Vision (ICCV) . Montreal, QC, Canada: IEEE, Oct. 2021, pp. 528–537. [Online]. Available: https://ieeexplore.ieee.org/document/9710410/

    work page arXiv 2021

  48. [48]

    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 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . Honolulu, HI: IEEE, Jul. 2017, pp. 2432–2443. [Online]. Available: https://ieeexplore.ieee.org/document/8099744/

    work page arXiv 2017

  49. [49]

    Joint 2D-3D-Semantic Data for Indoor Scene Understanding,

    I. Armeni, S. Sax, A. R. Zamir, and S. Savarese, “Joint 2D-3D-Semantic Data for Indoor Scene Understanding,” Apr. 2017, arXiv:1702.01105 [cs]. [Online]. Available: http://arxiv.org/abs/1702.01105

    Pith/arXiv arXiv 2017

  50. [50]

    3DMV: Joint 3D-Multi-View Prediction for 3D Semantic Scene Segmentation

    A. Dai and M. Nießner, “3DMV: Joint 3D-Multi-View Prediction for 3D Semantic Scene Segmentation,” Mar. 2018, arXiv:1803.10409 [cs]. [Online]. Available: http://arxiv.org/abs/1803.10409

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  51. [51]

    PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding

    S. Xie, J. Gu, D. Guo, C. R. Qi, L. J. Guibas, and O. Litany, “PointContrast: Unsupervised Pre-training for 3D Point Cloud Understanding,” Nov. 2020, arXiv:2007.10985 [cs]. [Online]. Available: http://arxiv.org/abs/2007.10985

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  52. [52]

    Pointly-supervised 3D Scene Parsing with Viewpoint Bottleneck

    L. Luo, B. Tian, H. Zhao, and G. Zhou, “Pointly-supervised 3D Scene Parsing with Viewpoint Bottleneck,” Sep. 2021, arXiv:2109.08553 [cs]. [Online]. Available: http://arxiv.org/abs/2109.08553

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  53. [53]

    Spconv: Spatially Sparse Convolution Library,

    S. Contributors, “Spconv: Spatially Sparse Convolution Library,” 2022. [Online]. Available: https://github.com/traveller59/spconv

    work page 2022

  54. [2018]

    Available: https://proceedings.neurips.cc/paper files/ paper/2018/hash/f5f8590cd58a54e94377e6ae2eded4d9-Abstract.html

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2018/hash/f5f8590cd58a54e94377e6ae2eded4d9-Abstract.html

    work page 2018

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.