Pith. sign in

REVIEW 4 major objections 5 minor 41 references

RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Two frozen foundation-model teachers, DINOv2 and CLIP, distilled into a semi-supervised segmentation student and fused into its decoder, deliver the best or near-best accuracy at 1–10% labels on three remote sensing benchmarks.

desk verdict The multi-teacher VFM distillation idea is reasonable and the code is there, but the FixMatch baseline inconsistency between tables undercuts the SOTA claim until the numbers are reconciled. read the letter →

arxiv 2506.08772 v2 pith:HELOHFBM submitted 2025-06-10 cs.CV

classification cs.CV
keywords semanticsegmentationsemi-supervisedlearningvisionfoundationmodelsremotesensingknowledgedistillationfeaturefusionDINOv2CLIP
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

RS-MTDF claims that semi-supervised semantic segmentation of remote sensing imagery—training a pixel classifier from a few labeled images plus many unlabeled ones—works better at very low label ratios when two frozen vision foundation models (DINOv2 and CLIP, large pretrained image encoders) act as auxiliary teachers. The paper is motivated by the observation that labeled and unlabeled remote sensing data often differ in distribution, and it argues that the teachers' robust features help the student generalize across that gap; the mechanism is feature-level distillation into the student encoder plus fusion of the aligned features back into the decoder. Across ISPRS Potsdam, LoveDA, and DeepGlobe at 1%, 5%, and 10% labels, the paper reports the best or near-best mean IoU among the compared methods, including the top LoveDA score at 1% labels (51.21% mIoU). Because all extra modules are used only during training, the reported gains come with no extra inference cost. If the claim holds, teams that can afford only tiny labeled sets for land-cover or urban mapping would get a stronger starting point than existing semi-supervised pipelines provide.

What carries the argument

The load-bearing mechanism is a translate–align–reproject–fuse loop around two frozen teacher encoders that stays active only during training. Feature translators are two-layer MLPs that map the student's final-layer features $f_S$ into each teacher's feature space, $\hat{f}_S^{(t)} = \mathrm{MLP}_t(f_S)$ for $t \in \{\mathrm{DINOv2}, \mathrm{CLIP}\}$; the distillation loss is the average over the two teachers of the mean squared error between the translated student features and the frozen teacher features. Linear projectors map the aligned features back to the student's channel dimension, and the decoder's top-level input is replaced by the weighted sum $f_{\mathrm{fused}} = \omega_S f_S + \omega_D(\tilde{f}_S^{\mathrm{DINO}} + \tilde{f}_S^{\mathrm{CLIP}})$, so the teachers' knowledge enters the final prediction as well as the encoder. The outer scaffold is FixMatch's mean teacher: an EMA-updated teacher produces confidence-thresholded pseudo-labels that supervise strongly augmented views of the same unlabeled images, and a DPT-style hierarchical decoder carries the fused features to the output. The mechanism's stated job is to keep the student anchored to stable, generalizable representations so the few labeled images do not dominate its learning.

What would settle it

Construct a deliberately distribution-shifted semi-supervised setup—for example, LoveDA with labeled images from one city and unlabeled images from the other two cities, or Potsdam split by acquisition season—and run RS-MTDF against FixMatch and UniMatchv2 under identical conditions. If the method's margin over these baselines does not grow relative to the paper's same-distribution random splits, and if the labeled-versus-unlabeled performance gap shown in the paper's Figure 1 does not shrink, the distribution-bridging explanation is not what the gains demonstrate.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the generalization of vision foundation models can be transferred into a semi-supervised segmentation student without fine-tuning the foundation models themselves. Two frozen encoders—DINOv2, chosen for spatially sensitive local features, and CLIP, chosen for global semantic priors from vision-language pretraining—act as expert teachers. The student's final encoder features are translated into each teacher's feature space by a dedicated two-layer MLP and aligned with the frozen teacher features by a mean-squared-error loss; the aligned features are then reprojected to the student's dimension and combined with the student's own high-level features into a weighted fused representation that replaces the decoder's top-level input. The paper reports that this pipeline matches or beats the compared methods on most settings across the three datasets, and its ablations show the DINOv2+CLIP pair outperforming either teacher alone and beating a CLIP+SAM pairing.

Load-bearing premise

The load-bearing premise is that the method's value comes from bridging a real distribution gap between the few labeled and the many unlabeled remote sensing images, yet no experiment in the paper tests that premise, because every labeled/unlabeled split is a random partition of one dataset and therefore shares one distribution.

Editorial extensions

If this is right

  • At 1% labels—the regime where annotation savings matter most—RS-MTDF reports the best mean IoU on all three datasets: 76.99% on Potsdam, 51.21% on LoveDA, and 67.15% on DeepGlobe.
  • The teacher-choice ablation on Potsdam at 1% supports the complementarity argument: DINOv2+CLIP reaches 76.99% mIoU versus 76.57% for DINOv2 alone and 75.74% for CLIP alone, and it also beats the CLIP+SAM pairing at 75.74%.
  • The DeepGlobe ablation at 10% shows the two modules are not interchangeable: adding distillation alone leaves mIoU essentially unchanged (72.12% versus a 72.13% baseline), while adding fusion on top jumps to 73.48%, so fusion carries the gain at higher supervision levels.
  • All added machinery—frozen teachers, translators, projectors—is confined to the training phase; inference uses only the student encoder–decoder, so the reported gains carry no deployment-time overhead.

Reading between the lines

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

  • The paper motivates the method as bridging a distribution gap between labeled and unlabeled remote sensing data, but no experiment creates such a gap: every split is random within one dataset, so labeled and unlabeled images are identically distributed; a cross-domain split (one city or season labeled, another unlabeled) would show whether the mechanism actually bridges the gap it names or whether
  • The fusion weights in $f_{\mathrm{fused}} = \omega_S f_S + \omega_D(\tilde{f}_S^{\mathrm{DINO}} + \tilde{f}_S^{\mathrm{CLIP}})$ are described only as 'determined empirically' and their values are never reported, so the balance between the student's own features and the teachers' contribution is not reproducible; publishing the sweep across weights and label ratios would make the fusion mechanism t
  • The 10% DeepGlobe pattern (distillation alone ≈ baseline, adding fusion +1.36 mIoU) suggests the two modules occupy different roles as supervision grows; testing distillation-only versus fusion-only across 1%, 5%, and 10% labels would show where each component is load-bearing.
  • Read strictly, the paper's own tables do not support 'consistently state-of-the-art' everywhere: at 10% labels on Potsdam and 5% on DeepGlobe, UniMatchv2 reports the higher mIoU, so the data are best summarized as best or near-best at most settings rather than first at every setting.
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

4 major / 5 minor

Summary. The paper proposes RS-MTDF, a semi-supervised semantic segmentation framework for remote sensing imagery. RS-MTDF builds on FixMatch and adds two frozen vision foundation models (DINOv2 and CLIP) as auxiliary teachers. The student encoder is distilled to match VFM feature spaces via learned translators, and the projected VFM-aligned features are fused into the student decoder. The method is evaluated on ISPRS Potsdam, LoveDA, and DeepGlobe at 1%, 5%, and 10% label ratios, and the authors claim consistent state-of-the-art performance, with ablations on teacher choice, confidence threshold, and the distillation/fusion modules.

Significance. The idea of using multiple frozen VFMs as feature-level teachers within a pseudo-labeling framework is plausible, and the paper provides a clear description of the losses, architecture, and ablations. The release of code is a practical strength. However, the central empirical claim is not yet supported: there is an unexplained 5.5 mIoU discrepancy for the FixMatch baseline between Table 1 and Table 4, there are no error bars or multiple-seed results despite several small margins, and the paper's motivating distribution-mismatch scenario is not tested because labeled and unlabeled sets are random splits of the same dataset. A directly relevant VFM-based baseline (SemiVL) is discussed but not included in the experiments. If the results survive a controlled re-evaluation, the method would be a useful contribution to remote sensing semi-supervised segmentation.

major comments (4)
  1. [Section 4.2.2, Tables 1 and 4] FixMatch on ISPRS Potsdam at the 1% label ratio is reported as 70.38 mIoU in Table 1 but as 75.90 mIoU in Table 4, under the same dataset and label ratio. The paper gives no explanation for this 5.52-point discrepancy, so the two rows were evidently produced under different conditions. This directly undermines the controlled comparison: the margin of the proposed method over FixMatch shrinks from +6.61 mIoU in Table 1 to +1.09 mIoU in Table 4. Please report the exact settings for each table, rerun all baselines under identical conditions, and clarify whether Table 4's FixMatch uses the same DINOv2 backbone and augmentation pipeline as RS-MTDF.
  2. [Section 4.2.2, Tables 1-3] No error bars, standard deviations, or multiple-seed results are reported, yet several claimed advantages are small (e.g., LoveDA 10%: 54.76 vs. 53.33 for UniMatchv2; DeepGlobe 5%: the proposed method at 71.70 is actually 0.13 mIoU below UniMatchv2 at 71.83). The statement in Section 4.2.2 that the method 'consistently achieves state-of-the-art performance' is therefore not statistically supported. Please provide mean and standard deviation over at least three seeds for the main comparison tables, and identify which differences are significant.
  3. [Section 1 and Section 4.1.1] The paper motivates RS-MTDF as addressing the 'distribution mismatch between limited labeled data and abundant unlabeled data' (Section 1, Figure 1), but in all experiments the labeled and unlabeled sets are random splits of a single dataset and therefore follow the same distribution. The proposed mechanism of bridging a real distribution gap with frozen VFM features is not tested by these experiments; the observed gains could come from generic regularization, feature alignment, or stronger initialization. Please add at least one experiment with a genuine distribution shift (e.g., training on one city and using another city as unlabeled data in LoveDA), or explicitly revise the mechanism claim to reflect the settings actually evaluated.
  4. [Section 4.2.1 and Section 2.3] SemiVL (Hoyer et al., 2024), a CLIP-based semi-supervised semantic segmentation method, is discussed in the related work but is not included in any comparison table. Since RS-MTDF's central contribution is the use of frozen VFM teachers, omitting this directly relevant baseline weakens the state-of-the-art claim. Please benchmark SemiVL (and ideally other VFM-guided SSS methods) under the same protocol as the other baselines.
minor comments (5)
  1. [Keywords and Section 3.2.1] There are several typos: 'semantic segmenattion' in the keywords, 'Archetecture' in Section 3.2.1, 'Visualizatiuon' in Section 4.2.3, 'Evaluation MMetrics' in Section 4.1.3, and 'valus' and 'grount truth' in the Figure 3 and Figure 4 captions.
  2. [Table 2] The table header refers to 'Kappa1' but the footnote for this marker is missing; please add the footnote or remove the superscript.
  3. [Section 4.1.2] The paper states that all experiments are run for 60 epochs but does not report the number of random seeds or the variance across runs; this is related to the major comment about error bars.
  4. [Algorithm 1 and Equation (6)] The notation for the projected student features, such as \hat{f}_S^{(t)}, is clear in context but would benefit from an explicit statement that the superscript (t) indexes the teacher model (DINOv2 or CLIP).
  5. [Figure 1] The caption of Figure 1 mentions 'mIoU performance on labeled and unlabeled samples', but the figure itself is not visible in the text; please ensure the figure is included and the caption describes the axes and curves.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical architecture/benchmark study; no prediction reduces by construction to a fitted input or a self-citation chain.

full rationale

RS-MTDF is an empirical semi-supervised segmentation paper with no formal derivation chain: the method is a concrete composition of FixMatch, frozen DINOv2/CLIP teachers, feature translators, and a weighted fusion module, all specified directly in Eqs. (5)-(9). The central claim, "consistently achieves state-of-the-art performance," is supported by benchmark comparisons against external methods (UniMatchv2, MUCA, DWL, AllSpark, etc.) on three external datasets, not by a quantity that is defined in terms of the claimed result. Loss weights, confidence thresholds, and fusion weights are ordinary hyperparameters, and the paper's ablation tables report internal component contributions rather than renaming a fitted parameter as a prediction. The visible inconsistency between FixMatch's mIoU in Table 1 (70.38) and Table 4 (75.90) under the same dataset and label ratio is a serious experimental-control concern, and the paper's distribution-mismatch motivation is not actually tested by its random splits, but neither issue is a circular-reasoning pattern. Minor self-citations (e.g., Li et al., 2024a) appear only in related-work motivation and do not carry the load of the empirical claim. No equation-level circularity, no self-citation chain, and no ansatz-smuggling is present, so the derivation is self-contained in the sense relevant to this review.

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

The central claim relies on the assumption that frozen VFM features provide a reliable training signal (Eqs. 5-8), on the inherited FixMatch consistency assumption (Eq. 2), and on several unstated or weakly justified choices: student initialization, strong augmentation for teachers, and empirically chosen fusion weights (Eq. 9).

free parameters (4)
  • Confidence threshold tau = 0.95
    Selected by ablation over 0.80, 0.85, 0.90, 0.95, 0.98 on Potsdam (Section 4.3.2); controls pseudo-label quality.
  • Loss weights lambda_l, lambda_u, lambda_d = 1/3 each
    Set to equal weighting in Section 4.1.2 without sensitivity analysis; the central results depend on these balances.
  • Fusion weights omega_S and omega_D = Not reported
    Introduced in Eq. 9 and described as 'determined empirically' (Section 3.2.3); these weights directly control how much VFM knowledge enters the decoder.
  • EMA momentum alpha = Not reported
    Used in Eq. 4 for teacher updates; value not specified in implementation details.
assumptions (4)
  • domain assumption Feature-level MSE alignment with frozen VFM representations transfers generalization to the student encoder.
    This is the core premise of the distillation loss in Eqs. 5-7 and Algorithm 1; the paper provides no analysis showing that matching DINOv2/CLIP features under strong augmentation is a valid training signal.
  • standard math The consistency regularization smoothness assumption inherited from FixMatch holds for remote sensing imagery.
    The framework builds directly on FixMatch (Section 3.1), which assumes predictions should be consistent under perturbations; whether this holds for remote sensing data is not separately validated.
  • ad hoc to paper The student encoder is initialized from DINOv2 pretrained weights.
    Section 4.1.2 says DINOv2-small is used as the encoder but does not state initialization; if the student starts from DINOv2 weights, the 'teacher' guidance is partially distilled from a larger sibling, which affects interpretation of the ablation results.
  • domain assumption Strongly augmented inputs yield meaningful DINOv2 and CLIP features for distillation.
    Algorithm 1 feeds the same strongly augmented image (with color jitter and CutMix) to the frozen teachers; the paper does not analyze the effect of strong augmentation on teacher feature quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation." pith.science (2026). https://pith.science/paper/HELOHFBM

@misc{pith2026250608772,
  author       = {Pith},
  title        = {Pith review of: RS-MTDF: Multi-Teacher Distillation and Fusion for Remote Sensing Semi-Supervised Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HELOHFBM}},
  note         = {Machine review of arXiv:2506.08772}
}
read the original abstract

Semantic segmentation in remote sensing images is crucial for various applications, yet its performance is heavily reliant on large-scale, high-quality pixel-wise annotations, which are notoriously expensive and time-consuming to acquire. Semi-supervised semantic segmentation (SSS) offers a promising alternative to mitigate this data dependency. However, existing SSS methods often struggle with the inherent distribution mismatch between limited labeled data and abundant unlabeled data, leading to suboptimal generalization. To alleviate this issue, we attempt to introduce the Vision Foundation Models (VFMs) pre-trained on vast and diverse datasets into the SSS task since VFMs possess robust generalization capabilities that can effectively bridge this distribution gap and provide strong semantic priors for SSS. Inspired by this, we introduce RS-MTDF (Multi-Teacher Distillation and Fusion), a novel framework that leverages the powerful semantic knowledge embedded in VFMs to guide semi-supervised learning in remote sensing. Specifically, RS-MTDF employs multiple frozen VFMs (e.g., DINOv2 and CLIP) as expert teachers, utilizing feature-level distillation to align student features with their robust representations. To further enhance discriminative power, the distilled knowledge is seamlessly fused into the student decoder. Extensive experiments on three challenging remote sensing datasets demonstrate that RS-MTDF consistently achieves state-of-the-art performance. Notably, our method outperforms existing approaches across various label ratios on LoveDA and secures the highest IoU in the majority of semantic categories. These results underscore the efficacy of multi-teacher VFM guidance in significantly enhancing both generalization and semantic understanding for remote sensing segmentation. Ablation studies further validate the contribution of each proposed module.

Figures

Figures reproduced from arXiv: 2506.08772 by the authors.

Figure 1
Figure 1. This figure illustrates the performance gap in mIoU between labeled and unlabeled data across FixMatch (Sohn et al., 2020), UniMatchv2 (Yang et al., 2025), and our pro￾posed method. Both FixMatch and UniMatchv2 exhibit a large discrepancy between the two data types, indicating limited generalization to unlabeled samples. In contrast, our method effectively reduces this gap while also achieving a notable improvement … view at source ↗
Figure 2
Figure 2. The overall architecture of our proposed RS-MTDF model. It consists of three main components: the student model (which learns from both labeled and unlabeled data), the teacher model (updated via EMA from the student for pseudo-labeling), and the frozen VFMs acting as auxiliary expert teachers. VFMs provide stable and rich prior knowledge through feature distillation and fusion. Fixmatch Unimatchv2 Ours GT Input Ima… view at source ↗
Figure 3
Figure 3. Qualitative comparison of different semi-supervised segmentation methods on the ISPRS Potsdam dataset. From left to right: input image, FixMatch, UniMatchv2, our method, fully supervised and grount truth. J, Song et al.: Preprint submitted to Elsevier Page 20 of 19 [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Ablation Study on different valus of confidence threshold in our method on the Potsdam dataset J, Song et al.: Preprint submitted to Elsevier Page 21 of 19 [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 30 canonical work pages

  1. [1]

    , author Ortego, D

    author Arazo, E. , author Ortego, D. , author Albert, P. , author O’Connor, N.E. , author McGuinness, K. , year 2020 . title Pseudo-labeling and confirmation bias in deep semi-supervised learning , in: booktitle 2020 International Joint Conference on Neural Networks (IJCNN) , pp. pages 1--8 . :10.1109/IJCNN48605.2020.9207304

  2. [2]

    , author Chen, D

    author Bai, Y. , author Chen, D. , author Li, Q. , author Shen, W. , author Wang, Y. , year 2023 . title Bidirectional copy-paste for semi-supervised medical image segmentation , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. pages 11514--11524

  3. [3]

    , author Touvron, H

    author Caron, M. , author Touvron, H. , author Misra, I. , author Jegou, H. , author Mairal, J. , author Bojanowski, P. , author Joulin, A. , year 2021 . title Emerging properties in self-supervised vision transformers , in: booktitle 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pp. pages 9630--9640 . :10.1109/ICCV48922.2021.00951

  4. [4]

    , author Papandreou, G

    author Chen, L.C. , author Papandreou, G. , author Kokkinos, I. , author Murphy, K. , author Yuille, A.L. , year 2017 . title Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs . https://arxiv.org/abs/1606.00915, arXiv:1606.00915 http://arxiv.org/abs/1606.00915

  5. [5]

    , author Yuan, Y

    author Chen, X. , author Yuan, Y. , author Zeng, G. , author Wang, J. , year 2021 . title Semi-supervised semantic segmentation with cross pseudo supervision , in: booktitle IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

  6. [6]

    , author Koperski, K

    author Demir, I. , author Koperski, K. , author Lindenbaum, D. , author Pang, G. , author Huang, J. , author Basu, S. , author Hughes, F. , author Tuia, D. , author Raskar, R. , year 2018 . title Deepglobe 2018: A challenge to parse the earth through satellite images , in: booktitle Proceedings of the IEEE conference on computer vision and pattern recogni...

  7. [7]

    , author Kukleva, A

    author Fan, Y. , author Kukleva, A. , author Dai, D. , author Schiele, B. , year 2023 . title Revisiting consistency regularization for semi-supervised learning . journal International Journal of Computer Vision volume 131 , pages 626--643

  8. [8]

    , author Liu, Y

    author Guo, Y. , author Liu, Y. , author Georgiou, T. , author Lew, M.S. , year 2018 . title A review of semantic segmentation using deep neural networks . journal International journal of multimedia information retrieval volume 7 , pages 87--93

Show all 41 references
  1. [9]

    , author Zhang, X

    author He, K. , author Zhang, X. , author Ren, S. , author Sun, J. , year 2016 . title Deep residual learning for image recognition , in: booktitle Proceedings of the IEEE conference on computer vision and pattern recognition , pp. pages 770--778

  2. [10]

    , author Zhou, Y

    author He, X. , author Zhou, Y. , author Zhao, J. , author Zhang, D. , author Yao, R. , author Xue, Y. , year 2022 a. title Swin transformer embedding unet for remote sensing image semantic segmentation . journal IEEE Transactions on Geoscience and Remote Sensing volume 60 , p...

  3. [11]

    , author Wang, J

    author He, Y. , author Wang, J. , author Liao, C. , author Shan, B. , author Zhou, X. , year 2022 b. title Classhyper: Classmix-based hybrid perturbations for deep semi-supervised semantic segmentation of remote sensing imagery . journal Remote Sensing volume 14 , pages 879

  4. [12]

    , author Kim, J

    author Heo, B. , author Kim, J. , author Yun, S. , author Park, H. , author Kwak, N. , author Choi, J.Y. , year 2019 . title A comprehensive overhaul of feature distillation , in: booktitle Proceedings of the IEEE/CVF international conference on computer vision , pp. pages 1921--1930

  5. [13]

    , author Tan, D.J

    author Hoyer, L. , author Tan, D.J. , author Naeem, M.F. , author Van Gool, L. , author Tombari, F. , year 2024 . title Semivl: semi-supervised semantic segmentation with vision-language guidance , in: booktitle European Conference on Computer Vision , organization Springer . ...

  6. [14]

    , author Shi, Y

    author Huang, W. , author Shi, Y. , author Xiong, Z. , author Zhu, X.X. , year 2024 . title Decouple and weight semi-supervised semantic segmentation of remote sensing images . journal ISPRS Journal of Photogrammetry and Remote Sensing volume 212 , pages 13--26 . https://www.s...

  7. [15]

    , author Cao, X

    author Li, K. , author Cao, X. , author Deng, Y. , author Song, J. , author Liu, J. , author Meng, D. , author Wang, Z. , year 2024 a. title Semicd-vl: Visual-language model guidance makes better semi-supervised change detector . journal IEEE Transactions on Geoscience and Rem...

  8. [16]

    , author Cao, X

    author Li, K. , author Cao, X. , author Meng, D. , year 2024 b. title A new learning paradigm for foundation model-based remote-sensing change detection . journal IEEE Transactions on Geoscience and Remote Sensing volume 62 , pages 1--12

  9. [17]

    , author Jiang, J

    author Li, K. , author Jiang, J. , author Codegoni, A. , author Han, C. , author Deng, Y. , author Chen, K. , author Zheng, Z. , author Chen, H. , author Liu, Z. , author Gu, Y. , et al., year 2024 c. title Open-cd: A comprehensive toolbox for change detection . journal arXiv ...

  10. [18]

    , author Liu, R

    author Li, K. , author Liu, R. , author Cao, X. , author Bai, X. , author Zhou, F. , author Meng, D. , author Wang, Z. , year 2025 . title Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images , in: booktitle Proceedings of the Computer Visi...

  11. [19]

    , author Hutter, F

    author Loshchilov, I. , author Hutter, F. , year 2019 . title Decoupled weight decay regularization . https://arxiv.org/abs/1711.05101, arXiv:1711.05101 http://arxiv.org/abs/1711.05101

  12. [20]

    , author Jiao, L

    author Lu, X. , author Jiao, L. , author Liu, F. , author Yang, S. , author Liu, X. , author Feng, Z. , author Li, L. , author Chen, P. , year 2022 . title Simple and efficient: A semisupervised learning framework for remote sensing image semantic segmentation . journal IEEE T...

  13. [21]

    , author Darcet, T

    author Oquab, M. , author Darcet, T. , author Moutakanni, T. , author Vo, H. , author Szafraniec, M. , author Khalidov, V. , author Fernandez, P. , author Haziza, D. , author Massa, F. , author El-Nouby, A. , author Assran, M. , author Ballas, N. , author Galuba, W. , author H...

  14. [22]

    , author Hudelot, C

    author Ouali, Y. , author Hudelot, C. , author Tami, M. , year 2020 . title Semi-supervised semantic segmentation with cross-consistency training , in: booktitle The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  15. [23]

    , author Mesejo, P

    author Peláez-Vegas, A. , author Mesejo, P. , author Luengo, J. , year 2023 . title A survey on semi-supervised semantic segmentation . https://arxiv.org/abs/2302.09899, arXiv:2302.09899 http://arxiv.org/abs/2302.09899

  16. [24]

    , author Kim, J.W

    author Radford, A. , author Kim, J.W. , author Hallacy, C. , author Ramesh, A. , author Goh, G. , author Agarwal, S. , author Sastry, G. , author Askell, A. , author Mishkin, P. , author Clark, J. , author Krueger, G. , author Sutskever, I. , year 2021 . title Learning transfe...

  17. [25]

    , author Bochkovskiy, A

    author Ranftl, R. , author Bochkovskiy, A. , author Koltun, V. , year 2021 . title Vision transformers for dense prediction , in: booktitle Proceedings of the IEEE/CVF international conference on computer vision , pp. pages 12179--12188

  18. [26]

    , author Lasinger, K

    author Ranftl, R. , author Lasinger, K. , author Hafner, D. , author Schindler, K. , author Koltun, V. , year 2020 . title Towardaw robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer . journal IEEE Transactions on Pattern Analysis and Machi...

  19. [27]

    , author Sohn, G

    author Rottensteiner, F. , author Sohn, G. , author Jung, J. , author Gerke, M. , author Baillard, C. , author Benitez, S. , author Breitkopf, U. , year 2012 . title The isprs benchmark on urban object classification and 3d building reconstruction . journal ISPRS Annals of the...

  20. [28]

    , author Berthelot, D

    author Sohn, K. , author Berthelot, D. , author Carlini, N. , author Zhang, Z. , author Zhang, H. , author Raffel, C.A. , author Cubuk, E.D. , author Kurakin, A. , author Li, C.L. , year 2020 . title Fixmatch: Simplifying semi-supervised learning with consistency and confidenc...

  21. [29]

    , author Yang, Y

    author Sun, B. , author Yang, Y. , author Zhang, L. , author Cheng, M.M. , author Hou, Q. , year 2024 . title Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation , in: booktitle Proceedings of the IEEE/CVF Conference on Computer Visi...

  22. [30]

    , author Mai, H

    author Sun, R. , author Mai, H. , author Zhang, T. , author Wu, F. , year 2023 . title Daw: exploring the better weighting function for semi-supervised semantic segmentation . journal Advances in Neural Information Processing Systems volume 36 , pages 61792--61805

  23. [31]

    , author Valpola, H

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

  24. [32]

    , author Zhang, Q

    author Wang, H. , author Zhang, Q. , author Li, Y. , author Li, X. , year 2024 . title Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation . https://arxiv.org/abs/2403.01818, arXiv:2403.01818 http://arxiv.org/abs/2403.01818

  25. [33]

    , author Zheng, Z

    author Wang, J. , author Zheng, Z. , author Lu, X. , author Zhong, Y. , year 2021 . title Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation , in: booktitle Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchma...

  26. [34]

    , author Sun, X

    author Wang, S. , author Sun, X. , author Chen, C. , author Hong, D. , author Han, J. , year 2025 . title Semi-supervised semantic segmentation for remote sensing images via multi-scale uncertainty consistency and cross-teacher-student attention . https://arxiv.org/abs/2501.10...

  27. [35]

    , author Shen, X

    author Wang, Y. , author Shen, X. , author Hu, S.X. , author Yuan, Y. , author Crowley, J.L. , author Vaufreydaz, D. , year 2022 . title Self-supervised transformers for unsupervised object discovery using normalized cut , in: booktitle Proceedings of the IEEE/CVF Conference o...

  28. [36]

    , author Sim \'e oni, O

    author Wysocza \'n ska, M. , author Sim \'e oni, O. , author Ramamonjisoa, M. , author Bursuc, A. , author Trzci \'n ski, T. , author P \'e rez, P. , year 2024 . title Clip-dinoiser: Teaching clip a few dino tricks for open-vocabulary semantic segmentation , in: booktitle Euro...

  29. [37]

    , author Qi, L

    author Yang, L. , author Qi, L. , author Feng, L. , author Zhang, W. , author Shi, Y. , year 2023 . title Revisiting weak-to-strong consistency in semi-supervised semantic segmentation , in: booktitle Proceedings of the IEEE/CVF conference on computer vision and pattern recogn...

  30. [38]

    , author Zhao, Z

    author Yang, L. , author Zhao, Z. , author Zhao, H. , year 2025 . title Unimatch v2: Pushing the limit of semi-supervised semantic segmentation . journal IEEE Transactions on Pattern Analysis and Machine Intelligence

  31. [39]

    , author Han, D

    author Yun, S. , author Han, D. , author Chun, S. , author Oh, S.J. , author Yoo, Y. , author Choe, J. , year 2019 . title Cutmix: Regularization strategy to train strong classifiers with localizable features , in: booktitle 2019 IEEE/CVF International Conference on Computer V...

  32. [40]

    , author Qi, X

    author Zhao, H. , author Qi, X. , author Shen, X. , author Shi, J. , author Jia, J. , year 2018 . title Icnet for real-time semantic segmentation on high-resolution images , in: booktitle ECCV

  33. [41]

    , author Yang, L

    author Zhao, Z. , author Yang, L. , author Long, S. , author Pi, J. , author Zhou, L. , author Wang, J. , year 2023 . title Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation , in: booktitle Proceedings of the IEEE/CVF Conference on ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.