Pith. sign in

REVIEW 3 major objections 6 minor 56 references

Source-Free Domain Adaptation for Geospatial Point Cloud Semantic Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read LoGo adapts a pretrained point-cloud segmentation model to a new sensor or city using only unlabeled target data, reaching 54.75% mIoU on the photogrammetry-to-UAV-LiDAR transfer and 73.54% on the ALS-to-MLS transfer.

desk verdict Plausible SFUDA method for geospatial point clouds, but headline mIoU numbers rest on test-set hyperparameter tuning; worth reviewing with major revisions. read the letter →

arxiv 2601.08375 v2 pith:FDZKPX6Z submitted 2026-01-13 cs.CV

Yuan Gao , Di Cao , Xiaohuan Xi , Sheng Nie , Shaobo Xia , Cheng Wang This is my paper · ORCID
classification cs.CV
keywords source-freedomainadaptationpointcloudsemanticsegmentationoptimaltransportpseudo-labelfilteringlong-taileddistributioncross-sensorshiftgeospatialremotesensingmean-teacherself-training
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that source-free unsupervised domain adaptation is practical for geospatial point-cloud semantic segmentation even when the shift is large—cross-scene and cross-sensor. It introduces LoGo, which uses only a pretrained model and unlabeled target data. The central claim is that local class-balanced prototype estimation and global optimal-transport alignment correct each other's biases, and that keeping only pseudo-labels where both views agree yields reliable supervision for self-training. If the claim holds, deployed segmentation models could be adapted to new regions or sensors without sharing raw source data, which matters where privacy rules or data regulations forbid it. The paper reports consistent gains over existing methods: mIoU of 54.75% on STPLS3D→H3D and 73.54% on DALES→T3D.

What carries the argument

Three interacting modules carry the argument. Class-balanced local prototype estimation (CBLPE) uses intra-class anchor mining—ranking confidence within each category and taking the top ρ fraction—to build robust prototypes for rare classes. Global distribution alignment (GDA) formulates pseudo-label assignment as an entropy-regularized optimal transport problem solved with the Sinkhorn-Knopp algorithm, enforcing a global class-prior marginal c_k estimated from the ensemble predictions. Local-global dual-consensus filtering (LGDCF) keeps only samples where the local ensemble prediction matches the global OT assignment. The framework runs as a mean-teacher with only Batch Normalization affine

What would settle it

Run LoGo on DALES→T3D with the OT class-prior marginal c set to three different vectors: the aggregate ensemble prediction, the true target class proportions from held-out labels, and a deliberately skewed vector (e.g., doubling Ground and halving Fence). If the mIoU swings by more than a few points across these settings, the global-prior assumption is load-bearing and the method's success depends on a prior that is itself derived from the biased model it is trying to correct.

Watch

Extended reading notes

Core claim

The paper's core discovery is a two-view supervision principle: a pseudo-label should be trusted only when local feature-similarity judgments and global distributional constraints agree. LoGo operationalizes this in three steps: class-balanced local prototype estimation mines the top-confidence anchors within each class rather than applying a global threshold, so tail classes keep usable prototypes; global distribution alignment solves an entropy-regularized optimal transport problem whose column marginal is the aggregate class prior of the target ensemble, preventing head classes from over-claiming ambiguous points; and a dual-consensus filter retains only points where the multi-augmented e

Load-bearing premise

The framework assumes that the class proportions computed from the pretrained model's aggregate predictions on the target data are close to the target's true class distribution, and it enforces those proportions as a hard constraint in optimal transport; if the estimate is biased, the global alignment pushes pseudo-labels toward the wrong class balance.

Editorial extensions

If this is right

  • Deployed segmentation models can be updated to a new city or sensor without access to source data, easing privacy and regulatory constraints.
  • Long-tailed classes such as fences, urban furniture, and utility lines receive usable supervision because prototypes are built per class rather than filtered by a global confidence threshold.
  • Optimal-transport global alignment prevents the winner-takes-all failure mode that pure prototype greedy assignment exhibits under severe domain shift.
  • The dual-consensus filter suppresses pseudo-label noise, which stabilizes self-training across epochs through the EMA teacher.

Reading between the lines

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

  • If the central claim holds, optimal-transport pseudo-label assignment is a natural template for other long-tailed 3D tasks where class imbalance and domain shift coincide, such as lidar segmentation for autonomous driving or building instance mapping.
  • The paper leaves implicit that the same global-prior mechanism could serve as a diagnostic: comparing the enforced OT marginal against an independent estimate of the target class distribution would reveal when the method's central assumption is breaking down.
  • A testable extension is to replace the ensemble-derived class prior with an external prior (for example, cadastral or land-cover statistics) and measure whether mIoU improves; that would separate the contribution of the prior estimate from the OT mechanism itself.
  • Because the method assumes closed-set classes, an open-set variant that allocates OT mass to a 'reject' category could handle target scenes with entirely new object types—a direction the paper names as future work.
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 / 6 minor

Summary. The paper introduces LoGo, a source-free unsupervised domain adaptation (SFUDA) framework for semantic segmentation of geospatial point clouds. The method freezes the source-pretrained feature extractor and classifier, updating only Batch Normalization parameters within a mean-teacher self-training loop. It combines three modules: class-balanced local prototype estimation with intra-class anchor mining, optimal-transport-based global distribution alignment with an estimated target class prior, and a local-global dual-consensus filter that retains only pseudo-labels where the ensemble prediction and the OT assignment agree. The method is evaluated on two transfer tasks, STPLS3D→H3D and DALES→T3D, reporting mIoU of 54.75% and 73.54%, respectively, which the paper claims as state-of-the-art under strict SFUDA.

Significance. If the reported results are reproducible under a valid SFUDA protocol, the paper makes a useful contribution. It addresses a practically important privacy-preserving setting, is among the first to target cross-sensor geospatial point cloud SFUDA, and the proposed combination of class-balanced prototype mining with OT-based global assignment is a sensible mechanism for long-tailed target distributions. The release of source code is a further strength, and the two benchmarks are well chosen. However, the current experimental validation has a load-bearing protocol issue: key hyperparameters are selected using target ground-truth labels, and no variance estimates are reported. These problems must be resolved before the state-of-the-art claim can be accepted.

major comments (3)
  1. [§4.2, §4.5.2, Fig. 9] Hyperparameters V (ensemble size) and ρ (anchor ratio) are selected from sensitivity curves computed against target test ground truth. The paper sets V=4, ρ=0.8 for H3D and V=4, ρ=0.7 for T3D at the peaks of these curves, with no held-out validation split and no unsupervised selection rule. This constitutes test-set tuning and violates the strict SFUDA protocol stated in §3.1.1. Because the final mIoU numbers are optimistically selected, the state-of-the-art comparison is unfair and the central empirical claim is not supported as stated. Please re-run with fixed hyperparameters, a validation scene, or a label-free criterion, and also report results at neighboring non-peak values to quantify sensitivity.
  2. [Tables 1–3] All quantitative results, including ablations, appear to be single runs with no error bars or significance tests. Self-training with stochastic multi-augmentation, EMA, and small target batch sizes is inherently variable, and several reported differences are small (e.g., CBLPE+GDA vs. full LoGo on STPLS3D: 52.75 vs. 54.75). Without at least three seeds with mean±std, or an equivalent statistical assessment, the claim that LoGo 'consistently outperforms' baseline methods is not established.
  3. [§3.3.2, Eq. (6)] The target class prior c is estimated from the source model's aggregate ensemble predictions and then enforced as a hard marginal in the optimal-transport assignment. The paper assumes that aggregate predictions provide a stable estimate of the true target class distribution, but under the severe sensor and viewpoint shifts considered here, this is precisely the quantity most likely to be biased. If c is wrong, OT enforces the wrong global proportions and the dual-consensus filter inherits that bias. This is a load-bearing modeling assumption. Please provide evidence for it, e.g., compare c to the true target proportions, add an oracle-prior ablation, or test robustness to perturbing c.
minor comments (6)
  1. [Abstract and Table 2] The abstract states an 11.42% mIoU improvement 'compared to the source-only baseline' on DALES→T3D, but Table 2 shows source-only mIoU of 56.09 and LoGo of 73.54, a difference of 17.45 points. The value 11.42 corresponds to the gain over Wang et al. (62.12). Please correct the comparison basis.
  2. [§2.3] The text contains an unresolved citation placeholder: 'Liu et al. [?]' after discussing BEV-projection methods. The reference list includes Liu et al. [47]; please ensure the in-text citation is properly linked and the surrounding claim is accurate.
  3. [§4.1.2] The two tasks define five common categories (Ground, Vehicle, Urban Furniture, Building, Vegetation), but the source and target datasets have 8–18 original labels. Please clarify explicitly how original labels are merged or mapped to these five categories, and whether the source model is trained only on the five mapped categories or on all original categories with the mapping applied only at evaluation.
  4. [§3.3.2, §4.2] Equation (6) includes an entropy regularization coefficient λ and the Sinkhorn algorithm, but the paper does not report λ, the number of Sinkhorn iterations, or the convergence tolerance. These are needed to reproduce the OT assignment exactly.
  5. [Fig. 9] The caption of Fig. 9 does not state that the sensitivity curves are computed against target ground-truth mIoU. Please state this explicitly and, ideally, show error bars over multiple runs so the reader can judge whether the selected peaks are stable.
  6. [§3.4, Eq. (8)] The dual-consensus filter retains only samples where the local prediction equals the Sinkhorn assignment. It would be useful to report the fraction of retained pseudo-labels per epoch/class, since a very low retention rate would indicate that the method is simply ignoring most of the target data.

Circularity Check

2 steps flagged · score 6.0 of 10

LoGo's OT 'global correction' is self-referential: the target class prior c is computed from the model's own ensemble predictions and then enforced as the OT marginal, so head-class over-dominance is imposed rather than corrected. Additionally, key hyperparameters V and ρ are selected from sensitivity curves evaluated against target test mIoU, making the reported SOTA numbers test-selected rather

  1. self definitional [Sec. 3.3.2, Eqs. (6)–(7)]
    "In our offline setting, we assume that while individual predictions may contain noise, the aggregate statistics provide a stable estimate of the domain’s class distribution. Thus, c is calculated from the global statistics of the ensemble predictions: c_k = |I_k| / Σ_j |I_j|. This transforms local statistical information into boundary conditions for global optimization."

    The GDA module is claimed to correct the over-dominance of head classes inherent in local greedy assignments, but its target marginal c is defined as the histogram of the very same local ensemble predictions (the I_k sets from Eq. (3)). Eq. (7) forces Q* to have column sums exactly equal to c, so the aggregate global pseudo-label distribution equals the aggregate local prediction histogram by construction. Thus head-class over-dominance is not corrected; it is imposed as a hard constraint. The paper's own wording admits that the 'boundary condition' for the global optimization is the local statistical information the method was supposed to rectify.

  2. fitted input called prediction [Sec. 4.2, Sec. 4.5.2, Fig. 9]
    "The mIoU peaks at ρ=0.8 for STPLS3D and ρ=0.7 for DALES. ... we adopt V=4 as the optimal trade-off between adaptation accuracy and computational cost."

    The sensitivity analysis in Fig. 9 plots mIoU against V and ρ, and mIoU requires target ground-truth labels. The final hyperparameters are then set to the peaks of these curves on the very benchmarks whose results are reported in Tables 1 and 2. No validation split or unsupervised selection rule is described. Consequently, the final LoGo numbers are generated with hyperparameters optimized against the target test labels, so the central SOTA claim is a test-selected result rather than an independent prediction under the strict source-free unsupervised protocol.

full rationale

The final performance claim is measured against held-out ground truth, so the paper is not fully circular, and there is no load-bearing self-citation chain or imported uniqueness theorem. However, two load-bearing steps reduce by construction. First, the OT module's global class prior c is computed from the model's own ensemble predictions and then enforced as the column marginal of Q*, so the 'global distribution alignment' cannot actually correct the global head-class bias; it only reassigns samples within the model's own predicted class proportions. Second, the reported benchmark numbers depend on V and ρ selected from sensitivity curves evaluated against target test mIoU on the same tasks, with no validation split; this makes the headline SOTA result a test-set selection rather than a protocol-compliant prediction. Because these issues affect the core contribution but the evaluation still involves an external ground-truth comparison, a score of 6 is appropriate: partial circularity, not total equivalence.

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

The central empirical claim rests on a self-training pipeline with several hand-set hyperparameters (ρ, V, λ, Sinkhorn iterations) and on the assumption that model-predicted class priors are reliable target priors for optimal transport. There are no invented physical entities or new theoretical objects; the prototypes and assignment plan are algorithmic constructs computed from data.

free parameters (4)
  • Anchor selection ratio ρ = 0.8 (STPLS3D→H3D), 0.7 (DALES→T3D)
    Chosen per dataset from sensitivity analysis on the target benchmark (Fig. 9b), i.e., tuned to maximize target test mIoU. Controls the fraction of each class used to build prototypes.
  • Ensemble size V = 4
    Selected from sensitivity analysis (Fig. 9a) as the peak mIoU on both target test sets; affects pseudo-label smoothing and computational cost.
  • Entropy regularization coefficient λ = not reported
    Appears in the Sinkhorn objective (Eq. 6) and controls the smoothness of the transport plan, but no value is given in Section 4.2 or the sensitivity analysis.
  • Sinkhorn iterations / convergence tolerance = not reported
    The iterative Sinkhorn-Knopp solver parameters are never stated; these affect the accuracy and speed of the optimal-transport assignment plan.
assumptions (5)
  • domain assumption Closed-set assumption: source and target label sets are identical (Y_S = Y_T).
    Section 3.1.1 explicitly assumes closed-set labels; the paper acknowledges future work is needed for unseen or missing categories. This is load-bearing because the prototype and OT machinery assume every target point belongs to one of the source classes.
  • ad hoc to paper Target class prior c, estimated from ensemble predictions, is a stable estimate of the true target class distribution.
    Section 3.3.2 states that 'aggregate statistics provide a stable estimate of the domain's class distribution' without validation. If the pretrained model's predictions are biased under domain shift, the OT marginals will encode those biases.
  • domain assumption Freezing feature extractor and classifier while updating only BN affine parameters is sufficient for target adaptation.
    Section 3.1.2 adopts BN-only fine-tuning to preserve source geometry while adapting statistics. This is a common but nontrivial hypothesis that is only tested on two tasks.
  • domain assumption Multi-augmented ensemble predictions on random rotations reduce variance enough to produce reliable anchors.
    Section 3.2.1 averages predictions over V random transformations; the exact augmentation set is not fully specified and V is tuned per dataset.
  • domain assumption Mean-teacher EMA with momentum α=0.999 stabilizes pseudo-label evolution.
    Section 3.5 relies on exponential moving average to prevent oscillation of pseudo-labels; this is a standard assumption from the semi-supervised learning literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Source-Free Domain Adaptation for Geospatial Point Cloud Semantic Segmentation." pith.science (2026). https://pith.science/paper/FDZKPX6Z

@misc{pith2026260108375,
  author       = {Pith},
  title        = {Pith review of: Source-Free Domain Adaptation for Geospatial Point Cloud Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDZKPX6Z}},
  note         = {Machine review of arXiv:2601.08375}
}
read the original abstract

Semantic segmentation of 3D geospatial point clouds is fundamental to remote sensing applications, yet domain shifts caused by regional and acquisition-related variations often degrade model performance. Although domain adaptation can mitigate such shifts, existing methods typically require access to source-domain data, which is often infeasible due to privacy concerns and regulatory policies. To address this, we propose LoGo (Local-Global Dual-Consensus), a novel source-free unsupervised domain adaptation (SFUDA) framework requiring only a pretrained model and unlabeled target data. At the local level, we introduce a class-balanced prototype estimation module that ensures that robust feature prototypes can be generated even for sample-scarce tail classes, effectively mitigating the feature collapse caused by long-tailed distributions. At the global level, we introduce an optimal transport-based global distribution alignment module that formulates pseudo-label assignment as a global optimization problem, effectively correcting the over-dominance of head classes inherent in local greedy assignments, and thereby preventing model predictions from being severely biased towards majority classes. Finally, we propose a dual-consistency pseudo-label filtering mechanism that retains only high-confidence pseudo-labels where local multi-augmented ensemble predictions align with global optimal transport assignments for self-training. Extensive experiments on two challenging benchmarks, encompassing cross-scene and cross-sensor settings, demonstrate that LoGo consistently outperforms existing state-of-the-art methods. The source code is available at https://github.com/GYproject/LoGo-SFUDA.

Figures

Figures reproduced from arXiv: 2601.08375 by the authors.

Figure 1
Figure 1. Overview of domain adaptation paradigms. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed LoGo framework. The architecture adheres to a parameter-e [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the domain adaptation scenario from photogrammetry-derived point clouds (STPLS3D) to UAV-based LiDAR point clouds (H3D). The top row compares the representative scenes, while the bottom row displays their corresponding semantic label spaces [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualization of the domain adaptation scenario from ALS point clouds (DALES) to MLS point clouds (T3D). The top row compares the represen￾tative scenes, while the bottom row displays their corresponding semantic label spaces. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of global semantic segmentation results on the H3D dataset. From left to right: Source-only, SHOT, TTYD, LoGo, and Ground [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of local semantic segmentation details on the H3D dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of global semantic segmentation results on the T3D dataset. From left to right: Source-only, SHOT, TTYD, LoGo, and Ground [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of local semantic segmentation details on the T3D dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Parameter sensitivity analysis on STPLS3D [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 2 linked inside Pith

  1. [1]

    S. Xia, S. Xu, R. Wang, J. Li, G. Wang, Building instance mapping from als point clouds aided by polygonal maps, IEEE Transactions on Geoscience and Remote Sensing 60 (2021) 1–13

    2021

  2. [2]

    Y . Gao, S. Xia, C. Wang, X. Xi, B. Yang, C. Xie, Se- mantic segmentation of airborne lidar point clouds with noisy labels, IEEE Transactions on Geoscience and Re- mote Sensing 62 (2024) 1–14

    2024

  3. [3]

    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 Geo- science and Remote Sensing 62 (2024) 1–12

    2024

  4. [4]

    C. R. Qi, L. Yi, H. Su, L. J. Guibas, Pointnet++: deep hierarchical feature learning on point sets in a metric space, in: Proceedings of the 31st International Confer- ence on Neural Information Processing Systems, Curran Associates Inc., 2017, p. 5105–5114

    2017

  5. [5]

    Thomas, C

    H. Thomas, C. R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, L. J. Guibas, Kpconv: Flexible and de- formable convolution for point clouds, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6411–6420. 14

    2019

  6. [6]

    H. Luo, K. Khoshelham, L. Fang, C. Chen, Unsupervised scene adaptation for semantic segmentation of urban mo- bile laser scanning point clouds, ISPRS Journal of Pho- togrammetry and Remote Sensing 169 (2020) 253–267

    2020

  7. [7]

    Michele, A

    B. Michele, A. Boulch, T.-H. Vu, G. Puy, R. Marlet, N. Courty, Train till you drop: Towards stable and robust source-free unsupervised 3d domain adaptation, in: Eu- ropean Conference on Computer Vision, Springer, 2024, pp. 1–19

    2024

  8. [8]

    Varney, V

    N. Varney, V . K. Asari, Q. Graehling, Dales: A large-scale aerial lidar data set for semantic segmentation, in: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 186–187

    2020

Show all 56 references
  1. [9]

    F. J. Lawin, M. Danelljan, P. Tosteberg, G. Bhat, F. S. Khan, M. Felsberg, Deep projective 3d semantic segmen- tation, in: International conference on computer analysis of images and patterns, Springer, 2017, pp. 95–107

    2017

  2. [10]

    Boulch, B

    A. Boulch, B. L. Saux, N. Audebert, Unstructured point cloud semantic labeling using deep segmentation net- works, in: Proceedings of the Workshop on 3D Object Retrieval, Eurographics Association, 2017, p. 17–24

    2017

  3. [11]

    Graham, M

    B. Graham, M. Engelcke, L. Van Der Maaten, 3d seman- tic segmentation with submanifold sparse convolutional networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 9224– 9232

    2018

  4. [12]

    C. Choy, J. Gwak, S. Savarese, 4d spatio-temporal con- vnets: Minkowski convolutional neural networks, in: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3075–3084

    2019

  5. [13]

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic graph cnn for learning on point clouds, ACM Trans. Graph. 38 (2019)

    2019

  6. [14]

    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 pat- tern recognition, 2020, pp. 11108–11117

    2020

  7. [15]

    H. Zhao, L. Jiang, J. Jia, P. H. Torr, V . Koltun, Point trans- former, in: Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 16259–16268

    2021

  8. [16]

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, H. Zhao, Point transformer v3: Sim- pler faster stronger, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 4840–4851

    2024

  9. [17]

    S. Zhao, Y . Wang, B. Li, B. Wu, Y . Gao, P. Xu, T. Darrell, K. Keutzer, epointda: An end-to-end simulation-to-real domain adaptation framework for lidar point cloud seg- mentation, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, 2021, pp. 3500–3509

    2021

  10. [18]

    L. Yi, B. Gong, T. Funkhouser, Complete & label: A domain adaptation approach to semantic segmentation of lidar point clouds, in: Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2021, pp. 15363–15373

    2021

  11. [19]

    A. Xiao, J. Huang, D. Guan, F. Zhan, S. Lu, Transfer learning from synthetic to real lidar point cloud for se- mantic segmentation, in: Proceedings of the AAAI con- ference on artificial intelligence, volume 36, 2022, pp. 2795–2803

    2022

  12. [20]

    G. Li, G. Kang, X. Wang, Y . Wei, Y . Yang, Adversarially masking synthetic to mimic real: Adaptive noise injection for point cloud segmentation adaptation, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20464–20474

    2023

  13. [21]

    Z. Yuan, M. Cheng, W. Zeng, Y . Su, W. Liu, S. Yu, C. Wang, Prototype-guided multitask adversarial network for cross-domain lidar point clouds semantic segmenta- tion, IEEE Transactions on Geoscience and Remote Sens- ing 61 (2023) 1–13

    2023

  14. [22]

    L. Kong, N. Quader, V . E. Liong, Conda: Unsupervised domain adaptation for lidar segmentation via regularized domain concatenation, 2023 IEEE International Confer- ence on Robotics and Automation (ICRA) (2021) 9338– 9345

    2023

  15. [23]

    Saltori, F

    C. Saltori, F. Galasso, G. Fiameni, N. Sebe, E. Ricci, F. Poiesi, Cosmix: Compositional semantic mix for do- main adaptation in 3d lidar segmentation, in: European conference on computer vision, Springer, 2022, pp. 586– 602

    2022

  16. [24]

    A. Xiao, J. Huang, D. Guan, K. Cui, S. Lu, L. Shao, Po- larmix: A general data augmentation technique for lidar point clouds, Advances in Neural Information Processing Systems 35 (2022) 11035–11048

    2022

  17. [25]

    Shaban, J

    A. Shaban, J. Lee, S. Jung, X. Meng, B. Boots, Lidar- uda: Self-ensembling through time for unsupervised li- dar domain adaptation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19784–19794

    2023

  18. [26]

    H. Luo, Z. Chen, F. Ye, T. Huang, H. He, W. Hu, Cross- sensor adaptive semantic segmentation for mobile laser scanning point clouds based on continuous potential scene surface reconstruction, ISPRS Journal of Photogramme- try and Remote Sensing 228 (2025) 537–551

    2025

  19. [27]

    J. N. Kundu, N. Venkat, R. V . Babu, et al., Univer- sal source-free domain adaptation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4544–4553

    2020

  20. [28]

    V . K. Kurmi, V . K. Subramanian, V . P. Namboodiri, Do- main impression: A source data free domain adaptation 15 method, in: Proceedings of the IEEE/CVF winter confer- ence on applications of computer vision, 2021, pp. 615– 625

    2021

  21. [29]

    Z. Qiu, Y . Zhang, H. Lin, S. Niu, Y . Liu, Q. Du, M. Tan, Source-free domain adaptation via avatar prototype gener- ation and adaptation, in: Z.-H. Zhou (Ed.), Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, International Join...

    2021

  22. [30]

    R. Li, Q. Jiao, W. Cao, H.-S. Wong, S. Wu, Model adap- tation: Unsupervised domain adaptation without source data, in: 2020 IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2020, pp. 9638– 9647

    2020

  23. [31]

    Y . Liu, W. Zhang, J. Wang, Source-free domain adap- tation for semantic segmentation, in: 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 1215–1224

    2021

  24. [32]

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, T. Dar- rell, Tent: Fully test-time adaptation by entropy mini- mization, in: International Conference on Learning Rep- resentations, 2021

    2021

  25. [33]

    M. J. Mirza, J. Micorek, H. Possegger, H. Bischof, The norm must go on: Dynamic unsupervised domain adapta- tion by normalization, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14765–14775

    2022

  26. [34]

    S. Yang, Y . Wang, J. Van De Weijer, L. Herranz, S. Jui, Generalized source-free domain adaptation, in: Proceed- ings of the IEEE/CVF international conference on com- puter vision, 2021, pp. 8978–8987

    2021

  27. [35]

    S. Yang, S. Jui, J. Van De Weijer, et al., Attracting and dis- persing: A simple approach for source-free domain adap- tation, Advances in Neural Information Processing Sys- tems 35 (2022) 5802–5815

    2022

  28. [36]

    Tanwisuth, X

    K. Tanwisuth, X. Fan, H. Zheng, S. Zhang, H. Zhang, B. Chen, M. Zhou, A prototype-oriented framework for unsupervised domain adaptation, Advances in Neural In- formation Processing Systems 34 (2021) 17194–17208

    2021

  29. [37]

    Karim, N

    N. Karim, N. C. Mithun, A. Rajvanshi, H.-p. Chiu, S. Samarasekera, N. Rahnavard, C-sfda: A curriculum learning aided self-training framework for efficient source free domain adaptation, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023...

    2023

  30. [38]

    Liang, D

    J. Liang, D. Hu, J. Feng, Do we really need to access the source data? source hypothesis transfer for unsuper- vised domain adaptation, in: International conference on machine learning, PMLR, 2020, pp. 6028–6039

    2020

  31. [39]

    J. Duan, Y . Zhang, Y . Qu, Source-free domain adaptation for point cloud semantic segmentation, in: 2024 IEEE In- ternational Conference on Multimedia and Expo (ICME), IEEE, 2024, pp. 1–6

    2024

  32. [40]

    Saltori, S

    C. Saltori, S. Lathuiliére, N. Sebe, E. Ricci, F. Galasso, Sf-uda 3d: Source-free unsupervised domain adaptation for lidar-based 3d object detection, in: 2020 International Conference on 3D Vision (3DV), IEEE, 2020, pp. 771– 780

    2020

  33. [41]

    Hegde, V

    D. Hegde, V . Kilic, V . Sindagi, A. B. Cooper, M. Fos- ter, V . M. Patel, Source-free unsupervised domain adap- tation for 3d object detection in adverse weather, in: 2023 IEEE International Conference on Robotics and Automa- tion (ICRA), IEEE, 2023, pp. 6973–6980

    2023

  34. [42]

    D. Tsai, J. S. Berrio, M. Shan, E. Nebot, S. Worrall, Ms3d++: Ensemble of experts for multi-source unsuper- vised domain adaptation in 3d object detection, IEEE Transactions on Intelligent Vehicles 10 (2025) 1999– 2014

    2025

  35. [43]

    Z. Xia, H. Xia, S. Jin, W. Wang, Z. Ding, X. Cao, Dspf: Dual-stage preservation and fusion for source-free do- main adaptive point cloud completion, in: Proceedings of the 33rd ACM International Conference on Multime- dia, 2025, pp. 8428–8437

    2025

  36. [44]

    X. He, Z. Zhu, L. Nan, H. Chen, J. Qin, M. Wei, Pointsfda: Source-free domain adaptation for point cloud comple- tion, arXiv preprint arXiv:2503.15144 (2025)

    2025 arXiv

  37. [45]

    S. Wang, Y . Tong, X. Shang, Z. Zhang, Multi-confidence guided source-free domain adaption method for point cloud primitive segmentation, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2024, pp. 737–743

    2024

  38. [46]

    S. Wang, Y . Tong, R. Xu, Z. Zhang, Sdaps*: A novel source-free domain adaption method for point cloud prim- itive segmentation, IEEE Transactions on Industrial Infor- matics 22 (2026) 14–25

    2026

  39. [47]

    Y . Liu, H. Zhu, Y . Lei, H. Liu, Y . Pei, Y . Guo, Sf-city: A source-free domain adaptation method for city-scale point cloud semantic segmentation, IEEE Transactions on Mul- timedia 27 (2025) 6908–6921

    2025

  40. [48]

    P. Wang, W. Yao, J. Shao, Z. He, Test-time adaptation for geospatial point cloud semantic segmentation with dis- tinct domain shifts, ISPRS Journal of Photogrammetry and Remote Sensing 229 (2025) 422–435

    2025

  41. [49]

    Tarvainen, H

    A. Tarvainen, H. Valpola, Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results, Advances in neural information processing systems 30 (2017)

    2017

  42. [50]

    Villani, et al., Optimal transport: old and new, volume 338, Springer, 2008

    C. Villani, et al., Optimal transport: old and new, volume 338, Springer, 2008. 16

    2008

  43. [51]

    Sinkhorn, Diagonal equivalence to matrices with pre- scribed row and column sums, The American Mathemat- ical Monthly 74 (1967) 402–405

    R. Sinkhorn, Diagonal equivalence to matrices with pre- scribed row and column sums, The American Mathemat- ical Monthly 74 (1967) 402–405

    1967

  44. [52]

    M. Chen, Q. Hu, Z. Yu, H. THOMAS, A. Feng, Y . Hou, K. McCullough, F. Ren, L. Soibelman, Stpls3d: A large- scale synthetic and real aerial photogrammetry 3d point cloud dataset, in: 33rd British Machine Vision Confer- ence 2022, BMVC 2022, London, UK, November 21-24, 2022, BMV...

    2022

  45. [53]

    Kölle, D

    M. Kölle, D. Laupheimer, S. Schmohl, N. Haala, F. Rot- tensteiner, J. D. Wegner, H. Ledoux, The hessigheim 3d (h3d) benchmark on semantic segmentation of high- resolution 3d point clouds and textured meshes from uav lidar and multi-view-stereo, ISPRS Open Journal of Pho- togra...

    2021

  46. [54]

    W. Tan, N. Qin, L. Ma, Y . Li, J. Du, G. Cai, K. Yang, J. Li, Toronto-3d: A large-scale mobile lidar dataset for seman- tic segmentation of urban roadways, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 2020, pp. 202–203

    2020

  47. [55]

    Y . Li, N. Wang, J. Shi, J. Liu, X. Hou, Revisiting batch normalization for practical domain adaptation, arXiv preprint arXiv:1603.04779 (2016)

    2016 arXiv

  48. [56]

    L. Yi, G. Xu, P. Xu, J. Li, R. Pu, C. Ling, I. McLeod, B. Wang, When source-free domain adaptation meets learning with noisy labels, in: The Eleventh International Conference on Learning Representations, 2023. 17

    2023

Pith tools

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