Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Resolving Primitive-Sharing Ambiguity in Long-Tailed TLS-Based Industrial MEP Point Cloud Segmentation via Spatial Context Constraints

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Spatial context constraints—neighborhood-entropy emphasis plus density normalization—can resolve primitive-sharing ambiguity that frequency-only re-weighting cannot, raising tail-class mIoU by 21.7% while holding head accuracy.

desk verdict A useful problem-framing paper with a simple loss tweak and a genuinely new reducer result; the entropy mechanism is under-verified, the reported numbers conflict, and the evaluation needs variance. read the letter →

arxiv 2601.19128 v2 pith:VTT3GPKH submitted 2026-01-27 cs.CV

Chao Yin , Qing Han , Zhiwei Hou , Yue Liu , Anjin Dai , Hongda Hu , Ji Yang , Wei Yao This is my paper · ORCID
classification cs.CV
keywords long-tailedpointcloudsegmentationprimitive-sharingambiguityclass-balancedlossneighborhoodentropyMEPcomponentsterrestriallaserscanningspatialcontextconstraintsdigitaltwin
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

The paper tries to establish that failure on rare but safety-critical industrial components is a geometric problem, not just a statistical one: tail classes like reducers and valves share cylindrical primitives with dominant pipes, so they are locally indistinguishable. It argues that frequency-based re-weighting alone cannot fix this, and introduces two loss-level spatial context constraints—Boundary-CB, which up-weights points whose neighbors disagree, and Density-CB, which normalizes scan density—that plug into any backbone. On the Industrial3D water-treatment dataset, Boundary-CB raises mIoU to 55.74% and tail-class mIoU by 21.7% relative while preserving head-class accuracy, turning reducer IoU from 0% to 21.12%. A sympathetic reader would care because these constraints offer a low-friction way to make safety-critical components recognizable in digital twins and scan-to-BIM workflows.

What carries the argument

Boundary-CB is the central mechanism: for each point, average the network's predicted class distribution over its k-nearest neighbors, compute Shannon entropy H_i, and weight that point's loss by (1+αH_i). This turns spatial disagreement into a per-point attention signal that highlights primitive-sharing boundaries without requiring new labels. Density-CB plays a supporting role: it modulates class-balanced weights by γ(d̄_c)=1/(1+log d̄_c), where d̄_c is the mean number of neighbors in a fixed radius for that class, countering scanner-distance density bias. Both are loss-level and architecture-agnostic.

What would settle it

During early training, record neighborhood entropy on reducer points that are misclassified as pipe. If those points have low entropy (so the modulator is close to 1), Boundary-CB's up-weighting is not acting on the points it claims to fix and the reducer gain must come from another mechanism.

Watch

Extended reading notes

Core claim

The central claim is that the 'dual crisis'—extreme class imbalance (215:1) compounded by geometric ambiguity from primitive sharing—cannot be resolved by loss re-weighting that only counts samples. The paper proposes neighborhood prediction consistency as the signal for ambiguity: Boundary-CB computes the entropy of the average prediction over k nearest neighbors and multiplies the class-balanced cross-entropy loss by (1+αH_i), focusing training on high-disagreement boundaries where reducers meet pipes or valve parts meet pipes. Density-CB applies a logarithmic inverse-density modulator to correct scan-distance bias. On Industrial3D with a point cloud segmentation backbone, the authors repo

Load-bearing premise

Boundary-CB assumes that the hard tail-class points are exactly the places where neighbor predictions disagree; if the model confidently labels a reducer as pipe, the entropy weight stays near one and the constraint cannot help.

Editorial extensions

If this is right

  • Any existing point cloud segmentation network can adopt the constraints by swapping the loss, so the improvement is additive to architectural advances.
  • Tail-class gains do not come at the expense of head classes: head mIoU rises slightly to 88.14%, contradicting the usual head-tail seesaw.
  • Safety-critical components like reducers and valves become detectable at usable IoU levels, removing a bottleneck for digital twin and scan-to-BIM knowledge extraction.
  • The optimal neighborhood size depends on component geometry (k=64 for linear pipe connections, k=32 for compact components), so tuning k per target object type is necessary.
  • Combining both constraints fails to beat either alone, suggesting the two factors are not independent in practice and a single geometry-aware modulator is preferable.

Reading between the lines

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

  • Editorial inference: The entropy-based mechanism could fail on tail classes that the early model is confidently wrong about; the paper does not measure how often reducer points are misclassified as pipe with low neighborhood entropy, so the claimed causal path is not directly evidenced.
  • Editorial inference: The 'dual crisis' framing predicts that any domain sharing geometric primitives between frequent and rare classes—for example, urban pole-vs-trunk or furniture-vs-wall—could see similar gains from neighborhood-consistency re-weighting, which is testable on existing benchmarks.
  • Editorial inference: Density-CB's small gain on Industrial3D suggests scan density is a minor factor relative to geometric ambiguity in this dataset; in mobile or sparse scanning scenarios the ranking may invert, which a controlled density-degradation experiment could confirm.
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 / 4 minor

Summary. The paper addresses long-tailed semantic segmentation of industrial MEP point clouds, identifying a 'dual crisis' of extreme class imbalance (215:1) compounded by geometric ambiguity from primitive sharing between head classes (Pipe, Duct, RectangularBeam) and tail classes (Valve, Reducer, Elbow, etc.). The authors propose two loss-level extensions of Class-Balanced (CB) Loss: Boundary-CB, which multiplies the CB weight by (1 + αH_i) where H_i is the entropy of the average predicted distribution over k nearest neighbors; and Density-CB, which multiplies the CB weight by a class-average density modulator γ(d) = 1/(1+log d). Both are architecture-agnostic and require only loss replacement. On the Industrial3D dataset with ResPointNet++, Boundary-CB is reported to achieve 55.74% mIoU (+1.65 pp over CB+Focal), tail mIoU 29.59% (+21.7% relative over the CE baseline), and Reducer IoU improving from 0% to 21.12%, while preserving head-class accuracy. The paper includes ablations on k, r, α, and β, qualitative results, and a discussion of limitations including Strainer remaining at 0% IoU.

Significance. If the reported gains are reproducible, the contribution is practically valuable: a simple, loss-level plug-in that improves rare safety-critical component segmentation without architectural changes. The problem formalization—distinguishing composite-tail from primitive-similarity tail classes—is well motivated and clearly presented. The paper also explicitly releases code, and the method has negligible training overhead, which are strengths. However, the empirical evidence currently rests on a single dataset and a single backbone with no error bars or multiple seeds, and the causal mechanism of Boundary-CB is not verified. The central claim that spatial context constraints resolve primitive-sharing ambiguity therefore remains plausible but unsupported by the evidence as presented.

major comments (3)
  1. [§3.4, Eqs. (8)–(10)] The causal mechanism of Boundary-CB is unverified and may not activate on the very points it claims to rescue. The modulation factor (1+αH_i) uses the entropy of the average predicted distribution over kNN. For a primitive-similarity tail point whose k neighbors all predict Pipe with high confidence—exactly the 'dual crisis' regime described in §1 and §3.1—H_i ≈ 0, so Boundary-CB degenerates to plain CB. The baseline confusion matrix in Fig. 8 shows Reducer is predicted as Pipe on 82.5% of its true points, so this is not a corner case. Yet the paper reports no distribution of H_i on true tail points, no epoch-wise trace, and no counterfactual such as random or ground-truth-boundary modulation. Without this, the +5.27 pp tail mIoU gain cannot be attributed to the proposed entropy-based spatial-context mechanism rather than to CB re-weighting plus an extra instance-level modulation. Please
  2. [§4.2, Table 2, Fig. 8 caption] There are internal numerical contradictions that undermine the experimental narrative. The text in §4.2 states that at k=64 'Valve declines by -5.3% IoU, Elbow by -4.5%, and Tee by -1.5%', but Table 2 shows Boundary-CB (k=64) improving all three classes relative to both the CE baseline (Valve 38.06→47.32, Elbow 40.77→43.23, Tee 3.10→3.24) and CB+Focal. Additionally, the Fig. 8 caption says Reducer's Pipe confusion decreases from 91% to 56%, while the confusion matrices show 82.5%→69.7%. Table 1 also contains a stray '.21' in the CB+Focal H-IoU cell. These inconsistencies must be resolved and the comparison basis for the 'Classes Decreased' discussion clarified.
  3. [§4.1.3, §4.4, Table 3] The empirical support for the central claim is narrow and hyperparameters are selected on the same benchmark. The paper evaluates only ResPointNet++ on Industrial3D, with no multiple seeds or error bars, while claiming the method is 'architecture-agnostic' and generalizable. The hyperparameters k=64, r=0.2, and α=1 are chosen on the same dataset used for the final reported numbers, so the headline performance may include selection bias. Please add at least one additional backbone or a second dataset, and/or report variance across seeds with the hyperparameter search explicitly separated from the final evaluation.
minor comments (4)
  1. [Abstract / Fig. 1 / §3.1] The imbalance ratio is inconsistently reported as 215:1 in the abstract and §3.1 but 203:1 in Fig. 1 and its caption. Please unify.
  2. [Fig. 7 caption] The legend uses 'LD-CB' and 'GC-CB', while the text and other figures use 'Density-CB' and 'Boundary-CB'. Please update for consistency.
  3. [Data Availability / Abstract] The abstract says code is available at a GitHub URL, but the Data Availability statement says code and trained models 'will be released upon acceptance'. Please clarify the current availability status.
  4. [§4.4/Table 4] Table 4 tests r = 0.2, 0.6, 0.8 m but the text implies smaller radii were also considered; please state the full search range or explain the selection.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the loss-level modulations are empirical and benchmark results are not reduced to fitted inputs or self-citations.

full rationale

The paper's core claims are empirical: Boundary-CB and Density-CB are loss-level reweighting schemes whose mIoU gains on Industrial3D are reported as experimental outcomes, not derived from the assumptions. The entropy term H_i in Eq. 9 is computed from network predictions and used as a loss weight (Eq. 10); this is a training-time input, not a quantity that the paper 'predicts' from its own construction. Hyperparameters k and alpha are selected by ablation on the same dataset, which is tuning rather than circular prediction. The backbone ResPointNet++ [7], CB Loss [19], and Focal Loss [33] are standard/external components; no uniqueness theorem or ansatz is imported via self-citation, and the 'dual crisis' framing is supported by dataset statistics and confusion-matrix analysis rather than by a self-referential proof. The skeptic's concern that confidently wrong predictions could yield low entropy is a legitimate robustness/correctness issue, but it does not constitute a circular step: the mechanism could fail empirically without the derivation being equivalent to its input.

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

No new physical entities are introduced. The central claim rests on one dataset's statistics and on the modeling assumption that prediction entropy reveals geometric ambiguity. Free hyperparameters (β, α, k, r) are fitted to maximize performance on the same benchmark, which is the main inflation risk.

free parameters (4)
  • beta (β) in CB loss = 0.9999
    Selected by sensitivity analysis (§4.4.2); sets the effective-number ceiling and thus all class weights.
  • Boundary-CB modulation strength α = 1.0
    Tuned in Figure 11d to maximize tail mIoU; controls how strongly neighborhood entropy is up-weighted.
  • Boundary-CB neighborhood size k = 64
    Ablated over {8,16,32,64,128}; k=64 gives best overall/tail mIoU on Industrial3D and is the central scale assumption.
  • Density-CB radius r = 0.2 m
    Ablated over {0.2, 0.6, 0.8}; 0.2 chosen as best; determines local density estimates.
assumptions (5)
  • ad hoc to paper Neighborhood prediction entropy H_i = -Σ_c p̂_c log p̂_c reliably indicates geometric ambiguity at primitive-sharing boundaries.
    Core mechanism in Eq. 9; if confidently-wrong regions have low entropy, the method down-weights exactly the hard tail points.
  • domain assumption Up-weighting high-entropy regions at the loss level improves tail-class IoU without degrading head classes.
    Training-dynamics assumption validated only on Industrial3D with one backbone; no second dataset or seed variance.
  • ad hoc to paper TLS density bias is adequately corrected by class-average density with γ(d)=1/(1+log d).
    Design choice in Eq. 7; no comparison with other density normalizations.
  • domain assumption Industrial3D's geometric analysis (86% of tail classes share cylindrical primitives with Pipe) is representative and correctly annotated.
    Used to define the dual crisis; dataset not yet public, so statistics cannot be independently checked.
  • domain assumption Class-Balanced effective-number formula captures diminishing information in 3D point clouds at point level.
    Borrowed from 2D vision [19] and assumed to apply to point-level counts and kNN neighborhoods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Resolving Primitive-Sharing Ambiguity in Long-Tailed TLS-Based Industrial MEP Point Cloud Segmentation via Spatial Context Constraints." pith.science (2026). https://pith.science/paper/VTT3GPKH

@misc{pith2026260119128,
  author       = {Pith},
  title        = {Pith review of: Resolving Primitive-Sharing Ambiguity in Long-Tailed TLS-Based Industrial MEP Point Cloud Segmentation via Spatial Context Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VTT3GPKH}},
  note         = {Machine review of arXiv:2601.19128}
}
read the original abstract

In terrestrial laser scanning (TLS)-based mechanical, electrical, and plumbing (MEP) point cloud segmentation, safety-critical components such as reducers and valves are persistently misclassifed, blocking reliable engineering knowledge extraction. This stems from a dual crisis--extreme class imbalance (215:1) compounded by geometric ambiguity, since most tail classes share cylindrical primitives with dominant head classes--that existing frequencybased re-weighting methods cannot resolve. We propose spatial context constraints that exploit neighborhood prediction consistency to disambiguate locally similar structures. Our approach extends Class-Balanced (CB) Loss with two architecture-agnostic mechanisms: Boundary-CB, an entropy-based constraint that emphasizes ambiguous boundaries and encodes an MEP assemblytopology prior, and Density-CB, a density-based constraint that compensates for scan-dependent variations and encodes TLS sensor-physics knowledge. Both operate at the loss level and integrate into existing pipelines without backbone modifcations. On the Industrial3D dataset (612.7M labelled points from water treatment facilities), our method achieves 55.74% mIoU, exceeding the strongest of three representative fully supervised backbone baselines (39.83-52.48% mIoU), with a 21.7% relative improvement on tail-class performance (29.59% vs. 24.32%) while preserving head-class accuracy (88.14%). Components with primitive-sharing ambiguity show strong gains: reducer improves from 0% to 21.12% IoU, and valve improves by 24.3% relative. These results show that spatial context constraints reduce primitive-sharing errors in the target industrial MEP setting and support more reliable identifcation of safety-critical components for Digital Twin and Scan-to-BIM applications. Code: https://github.com/PointCloudYC/LongTail3D.git.

Figures

Figures reproduced from arXiv: 2601.19128 by the authors.

Figure 1
Figure 1. Dataset Comparison Reveals Industrial3D’s Unique Dual Crisis. Pan￾els (a-b): Class distributions for S3DIS (indoor) and Industrial3D (water treatment facilities), showing head (blue), common (green), and tail (orange) classes sorted by fre￾quency. Panel (c): Quantitative comparison table revealing Industrial3D’s dual crisis: (1) Statistical severity: 3.5× more severe imbalance than S3DIS (215:1 vs. 62:1 ratio); (2) … view at source ↗
Figure 2
Figure 2. Spatial Context Constraints Framework for Resolving Primitive￾Sharing Ambiguity. Top (Pipeline): Our architecture-agnostic framework augments standard point cloud segmentation pipelines (encoder-decoder backbone networks like Res￾PointNet++ or RandLA-Net) with spatial context constraints. The input 3D point cloud undergoes hierarchical feature extraction and upsampling to produce per-point predic￾tions. Our constrai… view at source ↗
Figure 3
Figure 3. Ceiling Effect of Class-Balanced Loss. The effective number of samples Ec saturates as actual sample count nc increases, creating a ceiling effect that prevents extreme re-weighting. Different β values control the saturation rate: smaller β (e.g., 0.99) saturates quickly, while larger β (e.g., 0.9999) allows more gradual saturation. With β = 0.9999, classes beyond 10,000 samples receive nearly identical weights, sta… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Intuitive Framework for Understanding Spatial Context Constraints. This figure provides accessible conceptual understanding of how our three proposed meth￾ods resolve primitive-sharing ambiguity through intuitive metaphors. Panel A (CB Loss): Resource allocation metaph…
Figure 5
Figure 5. Figure 5: Mathematical Foundations and Mechanisms for Resolving Primitive￾Sharing Ambiguity. This figure formalizes how our spatial context constraints address geometric ambiguity through three complementary mechanisms. Top Row (Mathemat￾ical Concepts): Panel A shows the effecti…
Figure 6
Figure 6. Figure 6: Class Distribution in the Industrial3D Dataset. The 12 classes are grouped by frequency: Head classes (Duct, Pipe, RectangularBeam) comprise 77% of the 610M points; Common classes (Ibeam, Tank) have moderate representation; Tail classes (Flange, Elbow, etc.) each repre…
Figure 7
Figure 7. Figure 7: , Boundary-CB maintains lower validation loss throughout train￾ing and achieves faster convergence. 0 50 100 150 200 250 Epoch 0.0 0.5 1.0 1.5 2.0 Loss (a) Training and Validation Loss Baseline Baseline (Train) CB+Focal CB+Focal (Train) LD-CB LD-CB (Train) GC-CB GC-CB …
Figure 8
Figure 8. Figure 8: Confusion Matrices Reveal Resolution of Primitive-Sharing Ambigu￾ity. Left (Baseline): ResPointNet++ with standard cross-entropy loss shows significant confusion between geometrically similar classes, particularly Pipe dominating tail classes with cylindrical primitive…
Figure 9
Figure 9. Figure 9: Qualitative Segmentation Results on Industrial3D. Comparisons on three test areas showing ground truth (left), baseline predictions (center), and our predictions with spatial context constraints (right). Row 1 (Area 6-1): The baseline struggles with complex assemblies …
Figure 10
Figure 10. Figure 10: Component Ablation Study Validates Spatial Context Necessity. Progressive addition of constraint components starting from CE baseline, showing con￾tributions to overall mIoU (blue bars) and tail-class mIoU (red bars). Stages: (1) CE Baseline: standard cross-entropy ac…
Figure 11
Figure 11. Figure 11: Hyperparameter Sensitivity Analysis. The plots show the model’s sensi￾tivity to changes in the main hyperparameters: (a) β for CB, (b) radius r for Density￾CB, (c) neighbors k for Boundary-CB spatial context, and (d) modulation strength α for Boundary-CB. We plot both…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Industrial3D: A Water-Treatment TLS Point Cloud Dataset and Cross-Paradigm Benchmark for MEP Scene Understanding

    cs.CV 2026-03 accept novelty 6.5 of 10

    A 612M-point industrial MEP TLS dataset and cross-paradigm benchmark show best supervised mIoU of 55.74% versus 15.79% zero-shot Point-SAM, a 39.95-point domain gap from 215:1 imbalance and cylindrical ambiguity.

Reference graph

Works this paper leans on

51 extracted references · 7 canonical work pages · cited by 1 Pith paper

  1. [1]

    Mirzaei, M

    K. Mirzaei, M. Arashpour, E. Asadi, H. Masoumi, Y. Bai, A. Behnood, 3d point cloud data processing with machine learning for construction and infrastructure applications: A comprehensive review, Advanced En- gineering Informatics 51 (2022) 101501.doi:10.1016/j.aei.2021. 101501

    doi:10.1016/j.aei.2021 2022

  2. [2]

    Y. Guo, H. Wang, Q. Hu, H. Liu, L. Liu, M. Bennamoun, Deep learning for 3d point clouds: A survey, IEEE transactions on pattern analysis and machine intelligence 43 (12) (2020) 4338–4364.doi:10.1109/TPAMI. 2020.2977026

    arXiv 2020

  3. [3]

    Z. Ma, S. Liu, A review of 3d reconstruction techniques in civil engineer- ing and their applications, Advanced Engineering Informatics 37 (2018) 163–174.doi:10.1016/j.aei.2018.05.005

    work page doi:10.1016/j.aei.2018.05.005 2018

  4. [4]

    S. Chen, G. Fan, J. Li, Improving completeness and accuracy of 3d point clouds by using deep learning for applications of digital twins to civil structures, Advanced Engineering Informatics 58 (2023) 102196. doi:10.1016/j.aei.2023.102196

    arXiv 2023

  5. [5]

    J. Shao, W. Yao, P. Wang, Z. He, L. Luo, Urban geobim construction by integrating semantic lidar point clouds with as-designed bim models, IEEE Transactions on Geoscience and Remote Sensing 62 (2024) 1–12. doi:10.1109/TGRS.2024.3358370

    arXiv 2024

  6. [6]

    Agapaki, I

    E. Agapaki, I. Brilakis, Cloi-net: Class segmentation of industrial facil- ities’ point cloud datasets, Advanced Engineering Informatics 45 (2020) 101121.doi:10.1016/j.aei.2020.101121

    arXiv 2020

  7. [8]

    C. Yin, J. C. Cheng, B. Wang, V. J. Gan, Automated classification of piping components from 3d lidar point clouds using se-pseudogrid, Au- tomation in Construction 139 (2022) 104300.doi:10.1016/j.autcon. 2022.104300. 34

    arXiv 2022

  8. [9]

    H. Yue, Q. Wang, H. Huang, X. Xia, H. Fang, J. C. Cheng, Enhanc- ing semantic segmentation of mep scenes with deep learning and bim- generated synthetic point clouds, Advanced Engineering Informatics 68 (2025) 103723.doi:10.1016/j.aei.2025.103723

    arXiv 2025

Show all 51 references
  1. [10]

    H. Yue, Q. Wang, H. Zhao, N. Zeng, Y. Tan, Deep learning applications for point clouds in the construction industry, Automation in Construc- tion 168 (2024) 105769.doi:10.1016/j.autcon.2024.105769

    2024

  2. [11]

    Q. Wang, Automatic checks from 3d point cloud data for safety regula- tioncomplianceforscaffoldworkplatforms, AutomationinConstruction 104 (2019) 38–51.doi:10.1016/j.autcon.2019.04.008

    2019 doi

  3. [12]

    B. Wang, Q. Wang, J. C. Cheng, C. Song, C. Yin, Vision-assisted bim reconstruction from 3d lidar point clouds for mep scenes, Automa- tion in Construction 133 (2022) 103997.doi:10.1016/j.autcon.2021. 103997

    2022 doi

  4. [13]

    Armeni, O

    I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, S. Savarese, 3d semantic parsing of large-scale indoor spaces, in: Pro- ceedings of the IEEE conference on computer vision and pattern recog- nition, 2016, pp. 1534–1543.doi:10.1109/CVPR.2016.170

    2016 doi

  5. [14]

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, M. Nießner, Scannet: Richly-annotated 3d reconstructions of indoor scenes, in: Pro- ceedings of the IEEE conference on computer vision and pattern recog- nition, 2017, pp. 5828–5839.doi:10.1109/CVPR.2017.261

    2017 doi

  6. [15]

    C. R. Qi, H. Su, K. Mo, L. J. Guibas, Pointnet: Deep learning on point sets for 3d classification and segmentation, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 652– 660.doi:10.1109/CVPR.2017.16

    2017 doi

  7. [16]

    C. R. Qi, L. Yi, H. Su, L. J. Guibas, Pointnet++: Deep hierarchical feature learning on point sets in a metric space, in: Advances in neural information processing systems, Vol. 30, 2017, pp. 5099–5108.doi: 10.48550/arXiv.1706.02413

    2017 doi arXiv

  8. [17]

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y. Qiao, W. Ouyang, T. He, H.Zhao, Pointtransformerv3: Simpler, faster, stronger, in: Proceedings 35 of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 4840–4851.doi:10.1109/CVPR52733.2024.00463

    2024

  9. [18]

    Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, A. Markham, Randla-net: Efficient semantic segmentation of large- scale point clouds, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11108–11117.doi: 10.1109/CVPR426...

    2020

  10. [19]

    Y. Cui, M. Jia, T.-Y. Lin, Y. Song, S. Belongie, Class-balanced loss based on effective number of samples, in: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 9260– 9269.doi:10.1109/CVPR.2019.00949

    2019

  11. [20]

    Thomas, C

    H. Thomas, C. R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, L. J. Guibas, Kpconv: Flexible and deformable convolution for point clouds, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6411–6420.doi:10.1109/ICCV.2019.00651

    2019

  12. [21]

    Graham, L

    B. Graham, L. van der Maaten, Submanifold sparse convolutional net- works, arXiv preprint arXiv:1706.01307 (2017).doi:10.48550/arXiv. 1706.01307

    2017 doi arXiv

  13. [22]

    Kolodiazhnyi, A

    M. Kolodiazhnyi, A. Vorontsova, A. Konushin, D. Rukhovich, One- former3d: One transformer for unified point cloud segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20943–20953.doi:10.1109/CVPR52733. 2024.01979

    2024

  14. [24]

    Z. Liu, H. Hu, Y. Cao, Z. Zhang, X. Tong, A closer look at local aggregation operators in point cloud analysis, in: Computer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XXIII 16, Springer, 2020, pp. 326–342. doi:10.48550/arX...

    2020 doi arXiv

  15. [25]

    4398– 4407.doi:10.1109/cvpr52734.2025.00415

    G.Ji, S.Weder, F.Engelmann, M.Pollefeys, H.Blum, Arkitlabelmaker: A new scale for indoor 3d scene understanding, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 4398– 4407.doi:10.1109/cvpr52734.2025.00415

    2025

  16. [26]

    Zhang, B

    Y. Zhang, B. Kang, B. Hooi, S. Yan, J. Feng, Deep long-tailed learning: A survey, IEEE Transactions on Pattern Analysis and Machine Intelli- gence 45 (9) (2023) 10795–10816.doi:10.1109/TPAMI.2023.3268118

    2023

  17. [27]

    More, Survey of resampling techniques for improving classification performance in unbalanced datasets, arXiv preprint arXiv:1608.06048 (2016).doi:10.48550/arXiv.1608.06048

    A. More, Survey of resampling techniques for improving classification performance in unbalanced datasets, arXiv preprint arXiv:1608.06048 (2016).doi:10.48550/arXiv.1608.06048

    2016 doi arXiv

  18. [28]

    Zhang, M

    H. Zhang, M. Cisse, Y. N. Dauphin, D. Lopez-Paz, mixup: Beyond empirical risk minimization, arXiv preprint arXiv:1710.09412 (2017). doi:10.48550/arXiv.1710.09412

    2017 doi arXiv

  19. [29]

    Verma, A

    V. Verma, A. Lamb, C. Beckham, A. Najafi, I. Mitliagkas, D. Lopez- Paz, Y.Bengio, Manifoldmixup: Betterrepresentationsbyinterpolating hidden states, in: International conference on machine learning, PMLR, 2019, pp. 6438–6447

    2019

  20. [30]

    R. A. Rosu, P. Schütt, J. Quenzel, S. Behnke, Latticenet: Fast point cloud segmentation using permutohedral lattices (2020).arXiv:1912. 05905. URLhttps://arxiv.org/abs/1912.05905

    2020 arXiv

  21. [31]

    Cortinhal, G

    T. Cortinhal, G. Tzelepis, E. E. Aksoy, Salsanext: Fast, uncertainty- aware semantic segmentation of lidar point clouds for autonomous driv- ing (2024).arXiv:2003.03653. URLhttps://arxiv.org/abs/2003.03653

    2024 arXiv

  22. [32]

    Q. Hu, B. Yang, S. Khalid, W. Xiao, N. Trigoni, A. Markham, Sen- saturban: Learning semantics from urban-scale photogrammetric point clouds, International Journal of Computer Vision 130 (2) (2022) 316– 343.doi:10.1007/s11263-021-01554-9

    2022 doi

  23. [33]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollár, Focal loss for dense object detection, in: Proceedings of the IEEE international conference on computer vision, 2017, pp. 2980–2988.doi:10.1109/ICCV.2017. 324. 37

    2017 doi

  24. [34]

    Zhang, R

    X. Zhang, R. Xue, U. Soergel, A two-stage approach for rare class segmentation in large-scale urban point clouds, The International Archives of the Photogrammetry, Remote Sensing and Spatial In- formation Sciences XLIII-B2-2022 (2022) 329–334.doi:10.5194/ isprs-archives-XLIII-...

    2022

  25. [35]

    Zhang, D

    X. Zhang, D. Lin, R. Xue, U. Soergel, Target-guided learning for rare class segmentation in large-scale urban point clouds, The Inter- national Archives of the Photogrammetry, Remote Sensing and Spa- tial Information Sciences XLVIII-1/W2-2023 (2023) 1693–1698.doi: 10.5194/ispr...

    2023 doi

  26. [36]

    Lahoud, F

    J. Lahoud, F. S. Khan, H. Cholakkal, R. M. Anwer, S. Khan, Long- tailed 3d semantic segmentation with adaptive weight constraint and sampling, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 5037–5044.doi:10.1109/ICRA57147. 2024.10610029

    2024

  27. [37]

    M. Li, S. Lin, Z. Wang, Y. Shen, B. Zhang, L. Ma, Class-imbalanced semi-supervised learning for large-scale point cloud semantic segmen- tation via decoupling optimization, Pattern Recognition 156 (2024) 110701.doi:10.1016/j.patcog.2024.110701

    2024

  28. [38]

    Zhang, D

    X. Zhang, D. Lin, U. Soergel, Target-aware attentional network for rare class segmentation in large-scale lidar point clouds, ISPRS Journal of Photogrammetry and Remote Sensing 220 (2025) 32–50.doi:10.1016/ j.isprsjprs.2024.11.012

    2025

  29. [39]

    B. Zhu, Z. Jiang, X. Zhou, Z. Li, G. Yu, Class-balanced group- ing and sampling for point cloud 3d object detection, arXiv preprint arXiv:1908.09492 (2019).doi:10.48550/arXiv.1908.09492

    1908 doi arXiv

  30. [40]

    Cheng, X

    H. Cheng, X. Han, G.-S. Xiao, Spg: Structure-private graph network for 3d point cloud semantic segmentation with imbalanced classes, IEEE Transactions on Image Processing 32 (2023) 3887–3900

    2023

  31. [41]

    Huang, Y

    R. Huang, Y. Xu, D. Hong, W. Yao, P. Ghamisi, U. Stilla, Deep point embedding for urban classification using als point clouds: A new perspective from local to global, ISPRS Jour- nal of Photogrammetry and Remote Sensing 163 (2020) 62–81. 38 doi:https://doi.org/10.1016/j.isprsj...

    2020 doi

  32. [42]

    P. Wang, W. Yao, J. Shao, Z. He, Test-time adaptation for geospatial point cloud semantic segmentation with distinct domain shifts, ISPRS Journal of Photogrammetry and Remote Sensing 229 (2025) 422–435. doi:https://doi.org/10.1016/j.isprsjprs.2025.08.022. URLhttps://www.scienc...

    2025 doi

  33. [43]

    Hackel, N

    T. Hackel, N. Savinov, L. Ladicky, J. D. Wegner, K. Schindler, M. Polle- feys, Semantic3d. net: A new large-scale point cloud classification benchmark, arXiv preprint arXiv:1704.03847 (2017).doi:10.5194/ isprs-annals-iv-1-w1-91-2017

    2017 arXiv

  34. [44]

    M. Li, V. J. Gan, B. Wang, Integrating hierarchical segmentation and vision-language reasoning for spatially complex and occluded mep point clouds, Automation in Construction 179 (2025) 106455.doi:10.1016/ j.autcon.2025.106455

    2025

  35. [45]

    D. Hu, V. J. Gan, C. Yin, Robot-assisted mobile scanning for automated 3d reconstruction and point cloud semantic segmentation of building interiors, AutomationinConstruction152(2023)104949.doi:10.1016/ j.autcon.2023.104949

    2023

  36. [46]

    D. Hu, V. J. Gan, R. Zhai, Automated bim-to-scan point cloud seman- tic segmentation using a domain adaptation network with hybrid atten- tion and whitening (dawnet), Automation in Construction 164 (2024) 105473.doi:10.1016/j.autcon.2024.105473

    2024

  37. [47]

    C. Yin, B. Yang, J. C. Cheng, V. J. Gan, B. Wang, J. Yang, Label- efficient semantic segmentation of large-scale industrial point clouds us- ing weakly supervised learning, Automation in Construction 148 (2023) 104757.doi:10.1016/j.autcon.2023.104757

    2023

  38. [48]

    P. Wang, W. Yao, J. Shao, One class one click: Quasi scene-level weakly supervised point cloud semantic segmentation with active learning, IS- PRS Journal of Photogrammetry and Remote Sensing 204 (2023) 89– 104. 39

    2023

  39. [49]

    P. Wang, W. Yao, A new weakly supervised approach for als point cloud semantic segmentation, ISPRS Journal of Photogrammetry and Remote Sensing 188 (2022) 237–254.doi:10.1016/j.isprsjprs.2022.04.016

    2022 doi

  40. [50]

    H. Yue, Q. Wang, M. Zhang, Y. Xue, L. Lu, 2d–3d fusion approach for improved point cloud segmentation, Automation in Construction 177 (2025) 106336.doi:10.1016/j.autcon.2025.106336

    2025

  41. [51]

    Rozenberszki, O

    D. Rozenberszki, O. Litany, A. Dai, Language-grounded indoor 3d semantic segmentation in the wild, in: European Conference on Computer Vision, Springer, 2022, pp. 125–141.doi:10.1007/ 978-3-031-19827-4_8

    2022

  42. [52]

    Pierdicca, M

    R. Pierdicca, M. Paolanti, F. Matrone, M. Martini, C. Morbidoni, E. S. Malinverni, E. Frontoni, A. M. Lingua, Point cloud semantic segmen- tation using a deep learning framework for cultural heritage, Remote Sensing 12 (6) (2020) 1005.doi:10.3390/rs12061005

    2020 doi

  43. [53]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization (2017).arXiv:1412.6980,doi:10.48550/arXiv.1412.6980. URLhttps://arxiv.org/abs/1412.6980 40

    2017 doi arXiv

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.