Pith. sign in

REVIEW 4 major objections 4 minor 63 references

CoVar: Confidence-Variance-Guided Pseudo-Label Selection for Semi-Supervised Learning

T0 review · 4 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 Pseudo-label reliability needs both high confidence and low residual-class variance, not confidence alone.

desk verdict Useful plug-in heuristic with a broken theoretical derivation—the sign error in the CE expansion means the claimed entropy-minimization justification doesn't hold, but the empirical RCV signal is believable. read the letter →

arxiv 2601.11670 v3 pith:GYDTI7OC submitted 2026-01-16 cs.LG cs.AI

Jinshi Liu , Lei He , Pan Liu This is my paper
classification cs.LGcs.AI
keywords semi-supervisedlearningpseudo-labelselectionconfidencecalibrationresidual-classvarianceentropyminimizationspectralrelaxationsemanticsegmentationimageclassification
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 the standard practice of selecting pseudo-labels by a confidence threshold is fundamentally incomplete, because deep models are overconfident and high-confidence predictions can still be wrong. It derives, from entropy minimization, a second-order approximation of cross-entropy that separates a prediction's maximum confidence (MC) from the variance of probability mass over the non-maximum classes (RCV). The derivation shows that a reliable pseudo-label must have both high MC and low RCV, and that the penalty on RCV grows as confidence approaches 1, catching overconfident but unstable predictions. The paper then converts this two-dimensional criterion into a threshold-free selection rule using spectral relaxation in the MC-RCV plane, with Gaussian weighting to produce training weights, and reports consistent gains when plugged into existing segmentation and classification pipelines.

What carries the argument

The load-bearing object is the per-sample cross-entropy decomposition CE ≈ -log p_j(k') + [(K-1)^2/(2(1-p_j(k')))] v_j, obtained by Taylor-expanding log p_j(k) around the mean residual probability under the choice ε = μ_j. The residual-class variance (RCV) v_j measures how unevenly probability mass is spread across non-maximum classes. The confidence-dependent coefficient g_j(p_j(k')) = (K-1)^2/(2(1-p_j(k'))) makes the variance penalty grow as confidence approaches 1, which is what corrects overconfident but unstable predictions. This decomposition motivates the two-dimensional feature embedding [log p_j(k'), -((K-1)^2/(2(1-p_j(k')))) v_j], which is then partitioned by spectral relaxation (S

What would settle it

Measure the accuracy of pseudo-labels split by MC and RCV on a benchmark: if samples with high MC but high RCV are as accurate as those with high MC and low RCV, the central claim is wrong. A direct computational check is to compute, during training, the correlation between pseudo-label correctness and v_j within fixed confidence bins; the paper's criterion predicts a strong negative correlation inside high-confidence bins.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the cross-entropy used to train on pseudo-labels can be decomposed, to second order, as -log p_j(k') + [(K-1)^2 / (2(1-p_j(k')))] v_j, where p_j(k') is the maximum confidence and v_j is the variance of the remaining class probabilities around their mean. This decomposition implies that reliable predictions are exactly those that jointly maximize confidence and minimize residual-class variance, with an adaptive penalty that diverges as confidence goes to 1. The paper further claims that partitioning predictions in this confidence-variance plane by SVD-based spectral relaxation separates reliable from unreliable pseudo-labels without hand

Load-bearing premise

The whole reliability criterion depends on the choice to set the ideal distribution's residual-class probability ε equal to the model's own residual mean μ_j, so that 'high confidence plus low residual variance' follows algebraically from assuming the ideal residual allocation is uniform across non-maximum classes.

Editorial extensions

If this is right

  • Fixed confidence thresholds systematically over-select majority-class and overconfident-but-wrong pseudo-labels; jointly using MC and RCV reduces this bias.
  • The joint criterion can be added as a plug-in module to existing semi-supervised segmentation and classification methods with no inference-time overhead.
  • Gains are largest in low-label and class-imbalanced regimes, and even strong backbones continue to benefit from the added residual-dispersion check.
  • Replacing RCV with entropy, margin, or other scalar reliability metrics degrades performance, indicating that residual-class dispersion carries information these metrics miss.

Reading between the lines

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

  • The same cross-entropy decomposition could plausibly extend to other uncertainty-aware settings such as semi-supervised object detection or domain adaptation, though the paper only validates segmentation and classification.
  • Because the variance penalty diverges as confidence approaches 1, the method implicitly imposes a strict uniformity constraint on near-certain predictions; a testable question is whether this over-penalizes genuinely confident easy samples.
  • The spectral separation is performed within each batch; a temporal or online variant that accumulates statistics across epochs might stabilize selection further, especially in the early training phase.
  • The adaptive choice ε = μ_j embeds an assumption that the ideal residual allocation is uniform; on tasks with a skewed class structure, reweighting residual classes by their plausibility could improve the criterion further.
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 / 4 minor

Summary. The paper proposes CoVar, a pseudo-label selection framework for semi-supervised learning that jointly uses Maximum Confidence (MC) and Residual-Class Variance (RCV). The authors claim to derive this criterion from entropy minimization via a second-order Taylor expansion of cross-entropy, yielding an adaptive penalty that grows as confidence approaches one. They then cast pseudo-label selection as a spectral relaxation problem in the MC-RCV feature space and apply it to semantic segmentation (VOC, Cityscapes) and image classification (CIFAR-10, Mini-ImageNet) within existing SSL pipelines. Experiments report consistent gains over strong baselines, especially in low-label regimes.

Significance. If the theoretical derivation were sound, the paper would offer a principled, plug-in reliability module with no inference overhead, and the reported gains on DINOv2-B segmentation and Mini-ImageNet classification are notable. The manuscript also includes a limitations section and provides code. However, the central derivation contains a sign error that invalidates Eqs. (11) and (13) as approximations to cross-entropy, and the 'ideal distribution' choice introduces a circular element into the claimed entropy-minimization foundation. The qualitative conclusion (high MC + low RCV is good) survives a corrected derivation, but the paper's theoretical claims must be substantially repaired.

major comments (4)
  1. [Sec. III-B, Eq. (11)] The Taylor expansion sign is wrong. Substituting Eq. (10) into Eq. (9) gives CE = -log p + (K-1)ε log(p/μ) + (K-1)ε v/(2μ^2) + ... = -log p + (K-1)ε log(p/(1-p)) + ... , not -log p - (K-1)ε log(p/(1-p)). For K=2, p=0.9, ε=0.1, true CE = 0.325 but Eq. (11) returns -0.324. This equation is the foundation of the MC-RCV criterion, so the derivation as presented is invalid.
  2. [Sec. III-B, Eq. (13)] The term KL((1-p)/p) is not a well-defined KL divergence (the argument is a scalar) and is negative for p>0.5, so the claimed lower bound is wrong. With ε=μ_j, the correct simplification is CE ≈ -p log p - (1-p)log(1-p) + (K-1)^2/(2(1-p)) v, i.e., a binary entropy term plus the variance penalty. The paper's expression CE ≈ -log p + ... is not a lower bound and the subsequent batch-level decomposition inherits this error.
  3. [Sec. III-B.1] Setting ε=μ_j makes the ideal target distribution q depend on the model's own prediction. The statement 'reliable predictions have high MC and low RCV' then follows from the chosen surrogate (uniform residual allocation) rather than from an independent entropy-minimization principle. This is not necessarily fatal, but the paper should present ε=μ_j as a modeling choice and temper the claim that the criterion is 'derived from entropy minimization' without additional assumptions.
  4. [Sec. III-C and Algorithm 1] The abstract and Sec. III claim 'without hand-tuned confidence thresholds,' but the method uses hyperparameters λ=0.25 in the cluster-selection score and τ_c=0.95 in the classification setup. The paper should clarify which parts are truly threshold-free and which rely on manually chosen constants; otherwise the 'threshold-free' claim is overstated.
minor comments (4)
  1. [Eq. (13)] The KL term appears outside the summation but depends on j; the notation is ambiguous. If it is meant to be summed, parentheses are missing.
  2. [Supplementary Lemma 1] The remainder bound C'_j = (K-1)^{3/2}/(3(1-ρ)^3 μ_j^2) diverges as p_j(k')→1 (μ_j→0), so the 'uniform cubic bound' is not uniform over the range of interest. This should be stated explicitly.
  3. [Fig. 1] The caption text seems to describe the axes inconsistently ('vertical axis represents the mean of the maximum confidence' but the horizontal axis is 'computed using the validation set'). Please clarify.
  4. [General] Several citations (e.g., [1], [8]) are from the authors' own prior work; please ensure that the novelty relative to these works is clear, especially regarding the theoretical contribution.

Circularity Check

1 steps flagged · score 6.0 of 10

CoVar's MC-RCV criterion is not derived from entropy minimization; it is built in by setting ε=μ_j, so the central theoretical claim reduces to the chosen surrogate.

  1. self definitional [Sec. III-B.1 ('Setting the Ideal Residual-Class Confidence ε'), Eq. (13); cf. Eq. (8) and Supplementary Corollary 1]
    "Since concrete numerical values must be used in computation and ε would otherwise vary across samples and classes, we set ε in the ideal distribution q equal to the residual-class mean of the model prediction itself, i.e., ε = µj = 1−pj(k′) K−1 . Combined with Eq. 12, this yields the simplified form adopted in Eq. 13."

    The paper claims to derive 'high MC and low RCV' from entropy minimization, but the ideal target distribution q in Eq. (8) is defined with residual mass ε, and Sec. III-B.1 then sets ε equal to the model's own residual-class mean μ_j. The surviving cross-entropy terms are therefore -log p_j(k') plus a variance term, with the total residual-mass discrepancy removed by construction. If the paper had kept the actual entropy-minimization target (ε→0), the variance term would vanish and Eq. (11) would reduce to -log p_j(k'). Thus the central reliability criterion is the algebraic image of the ε=μ_j choice, not an independent consequence of entropy minimization. The supplementary 'Corollary 1' repeats the same substitution as if it were a proof.

full rationale

The experimental comparisons (PASCAL VOC, Cityscapes, CIFAR-10, Mini-ImageNet, etc.) are genuine external benchmarks and are not circular; CoVar's gains are evaluated against published baselines, and no load-bearing self-citation chain was found (CSL [8] is a baseline, not a premise). However, the theoretical derivation in Sec. III-B is partially circular in a specific, quotable way: the 'ideal' distribution q is built from the model's own residual-class mean. Eq. (13) sets ε=μ_j, so the decomposition CE ≈ -log p + [(K-1)^2/(2(1-p))]v is a consequence of that construction, and the conclusion that reliable pseudo-labels require low RCV follows by construction rather than from an independent first-principles argument. The spectral-relaxation module inherits this same construction through the embedding in Eq. (16), although the ablations (Tables V-VII) provide independent empirical evidence for the usefulness of the resulting weight. Separately, the sign of the middle term in Eqs. (11)/(13) appears to be incorrect, so the claimed CE approximation is invalid; that is a correctness risk rather than an additional circularity. Overall: central theory partially reduces to its own surrogate; score 6.

Assumptions & free parameters 2 free parameters · 5 assumptions · 1 invented entities

The central derivation rests on the self-referential choice ε=μ_j (Sec. III-B.1) and on Assumption 1 bounding residual probabilities around their mean (Supplementary Lemma 1). The 2-cluster spectral relaxation assumes separability of reliable and unreliable predictions in the MC-RCV plane. The only hand-fitted method-specific number is λ=0.25 used to choose the reliable cluster; the τ_c=0.95 threshold in the classification setup also contradicts the threshold-free claim.

free parameters (2)
  • λ (cluster-selection score weight) = 0.25
    Algorithm 1 line 18: score_c = μ_c[0] − λ σ_c[1] is used to choose the reliable cluster; λ is hand-set and no sensitivity sweep is reported.
  • τ_c (confidence threshold in classification setup) = 0.95
    Sec. IV-A: 'the confidence threshold was fixed at τ_c = 0.95' for the classification pipelines; this contradicts the claimed threshold-free selection if applied to CoVar.
assumptions (5)
  • ad hoc to paper Residual-scale boundedness: |δ_j(k)| ≤ ρ μ_j for all non-maximum classes (Supplementary Assumption 1).
    Necessary for the Taylor remainder bound in Lemma 1. Not empirically validated; if residual probabilities are highly uneven, the CE decomposition and RCV criterion lose their stated guarantees.
  • ad hoc to paper The ideal target distribution q has ε = μ_j, the model's own residual-class mean (Sec. III-B.1, Eq. 13).
    This makes the 'ideal' distribution depend on the prediction itself. The resulting high-MC/low-RCV criterion is therefore built into the chosen surrogate rather than derived from an independent entropy-minimization target.
  • ad hoc to paper Uniform allocation of residual-class probability is the ideal allocation.
    The paper defines low RCV as reliable because q is uniform on the non-maximum classes. No independent justification is given for why uniform residuals are the right reliability target.
  • domain assumption Predictions are separable into two reliable/unreliable clusters in the MC-RCV feature space.
    The spectral relaxation partitions into exactly two clusters; if the 2D feature distribution is not separable or is multimodal, the assignment rule in Eq. (18) can fail.
  • domain assumption Entropy minimization is a valid SSL principle.
    The derivation starts from entropy minimization, a standard SSL regularizer; the paper does not justify this principle beyond citing general SSL practice.
invented entities (1)
  • Residual-Class Variance (RCV) v_j
    purpose: Second-order statistic measuring dispersion of probability mass over non-maximum classes; used as the reliability feature paired with maximum confidence.
    RCV is defined entirely from the model's own predictive distribution; its usefulness is evidenced only by in-paper ablations, with no external falsifiable prediction or independent benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoVar: Confidence-Variance-Guided Pseudo-Label Selection for Semi-Supervised Learning." pith.science (2026). https://pith.science/paper/GYDTI7OC

@misc{pith2026260111670,
  author       = {Pith},
  title        = {Pith review of: CoVar: Confidence-Variance-Guided Pseudo-Label Selection for Semi-Supervised Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GYDTI7OC}},
  note         = {Machine review of arXiv:2601.11670}
}
read the original abstract

Pseudo-label selection in semi-supervised learning is commonly driven by maximum-confidence thresholds, yet confidence alone can be unreliable under model overconfidence and class imbalance. We propose CoVar, a confidence--variance framework that assesses pseudo-label reliability by jointly modeling Maximum Confidence (MC) and Residual-Class Variance (RCV). Starting from entropy minimization, we derive a second-order cross-entropy approximation showing that low-loss pseudo-labels are favored when MC is high and RCV is low, with a confidence-dependent penalty that becomes stronger for near-certain predictions. Based on this criterion, CoVar embeds predictions into a two-dimensional confidence--variance space and uses SVD-based spectral relaxation to separate reliable and unreliable predictions without hand-tuned confidence thresholds. Cluster-wise Gaussian weighting then converts this separation into per-sample training weights. The resulting weights can be integrated into existing semi-supervised segmentation and classification pipelines during training and introduce no inference-time overhead. Experiments on PASCAL VOC 2012, Cityscapes, CIFAR-10, CIFAR-100, SVHN, and STL-10 show clear gains on VOC and Cityscapes under matched backbones, as well as competitive or improved error rates on standard classification benchmarks. These results indicate that residual-class dispersion provides a useful signal complementary to confidence for robust pseudo-label selection.

Figures

Figures reproduced from arXiv: 2601.11670 by the authors.

Figure 1
Figure 1. Relationship between pseudo-label accuracy and maximum confidence [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Pseudo-label accuracy and selection ratio under different confidence [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall pipeline of the proposed method. Based on the proposed confidence-variance theory, the prediction separation module divides pseudo [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Illustration of the approximate CE loss CE = − log pj (k ′ )  + (K−1)2 2(1−pj (k′)) vj on the plane spanned by model confidence pj (k ′ ) and relative confidence variation vj . The contour map shows that (i) samples with low confidence and large variation (top-left re…
Figure 5
Figure 5. Figure 5: Change in pseudo-label selection rate during different model training [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The variation of the sample’s MC and RCV during model training. Experiments are conducted on the blender PASCAL VOC 2012 dataset (1/4, [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Ablation studies on different designs of CoVar. Experiments are [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Class-wise pseudo-label selection rates during training on blended [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 1
Figure 1. Figure 1: Evolution of pseudo-label selection during training on PASCAL VOC 2012 (1/8, 513 [PITH_FULL_IMAGE:figures/full_fig_p015_1.png]
Figure 2
Figure 2. Figure 2: Qualitative comparison of segmentation results on PASCAL VOC 2012 (1/8, 513 [PITH_FULL_IMAGE:figures/full_fig_p016_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 1 linked inside Pith

  1. [1]

    L. Yang, Z. Zhao, and H. Zhao, ``Unimatch v2: Pushing the limit of semi-supervised semantic segmentation,'' IEEE TPAMI, 2025

    2025

  2. [2]

    J. Wu, Z. Li, W. Sun, J. Yin, L. Nie, and Z. Lin, ``Clusmatch: Improving deep clustering by unified positive and negative pseudo-label learning,'' IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  3. [3]

    Tarubinga and J

    E. Tarubinga and J. Kalafatovich Espinoza, ``Confidence-weighted boundary-aware learning for semi-supervised semantic segmentation,'' arXiv e-prints, pp. arXiv--2502, 2025

    2025

  4. [4]

    H. Li, N. Wang, X. Yang, X. Wang, and X. Gao, ``An enhanced adaptive confidence margin for semi-supervised facial expression recognition,'' IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  5. [5]

    Q. Xie, Z. Dai, E. Hovy, T. Luong, and Q. Le, ``Unsupervised data augmentation for consistency training,'' NeurIPS, vol. 33, pp. 6256--6268, 2020

    2020

  6. [6]

    J. Kim, Y. Min, D. Kim, G. Lee, J. Seo, K. Ryoo, and S. Kim, ``Conmatch: Semi-supervised learning with confidence-guided consistency regularization,'' in ECCV. 1em plus 0.5em minus 0.4em Springer, 2022, pp. 674--690

    2022

  7. [7]

    Y. Wang, H. Chen, Y. Fan, W. Sun, R. Tao, W. Hou, R. Wang, L. Yang, Z. Zhou, L.-Z. Guo et al., ``Usb: A unified semi-supervised learning benchmark for classification,'' NeurIPS, vol. 35, pp. 3938--3961, 2022

    2022

  8. [8]

    Liu and J

    P. Liu and J. Liu, ``When confidence fails: Revisiting pseudo-label selection in semi-supervised semantic segmentation,'' in ICCV, October 2025, pp. 21\,874--21\,884

    2025

Show all 63 references
  1. [9]

    Xie, M.-T

    Q. Xie, M.-T. Luong, E. Hovy, and Q. V. Le, ``Self-training with noisy student improves imagenet classification,'' in CVPR, 2020, pp. 10\,687--10\,698

    2020

  2. [10]

    B. Zoph, G. Ghiasi, T.-Y. Lin, Y. Cui, H. Liu, E. D. Cubuk, and Q. Le, ``Rethinking pre-training and self-training,'' NeurIPS, vol. 33, pp. 3833--3845, 2020

    2020

  3. [11]

    J. Liu, Z. Jiang, W. Gui, Z. Chen, and C. Zhang, ``Occlusion segmentation: Restore and segment invisible areas for particle objects,'' IEEE TASE, 2024

    2024

  4. [12]

    Tarvainen and H

    A. Tarvainen and H. Valpola, ``Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results,'' NeurIPS, vol. 30, 2017

    2017

  5. [13]

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, ``Fixmatch: Simplifying semi-supervised learning with consistency and confidence,'' NeurIPS, vol. 33, pp. 596--608, 2020

    2020

  6. [14]

    Zhang, Y

    B. Zhang, Y. Wang, W. Hou, H. Wu, J. Wang, M. Okumura, and T. Shinozaki, ``Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,'' NeurIPS, vol. 34, pp. 18\,408--18\,419, 2021

    2021

  7. [15]

    Y. Wang, H. Chen, Q. Heng, W. Hou, Y. Fan, Z. Wu, J. Wang, M. Savvides, T. Shinozaki, B. Raj et al., ``Freematch: Self-adaptive thresholding for semi-supervised learning,'' in ICLR, 2023

    2023

  8. [16]

    H. Chen, R. Tao, Y. Fan, Y. Wang, J. Wang, B. Schiele, X. Xie, B. Raj, and M. Savvides, ``Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning,'' in ICLR, 2023

    2023

  9. [17]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, ``On calibration of modern neural networks,'' in Proceedings of the 34th International Conference on Machine Learning (ICML). 1em plus 0.5em minus 0.4em PMLR, 2017, pp. 1321--1330

    2017

  10. [18]

    J. C. Platt, ``Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods,'' Advances in large margin classifiers, vol. 10, no. 3, pp. 61--74, 1999

    1999

  11. [19]

    Zadrozny and C

    B. Zadrozny and C. Elkan, ``Obtaining calibrated probability estimates from decision trees and naive bayesian classifiers,'' in Proceedings of the 18th International Conference on Machine Learning (ICML). 1em plus 0.5em minus 0.4em Morgan Kaufmann, 2001, pp. 609--616

    2001

  12. [20]

    Szegedy, V

    C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, ``Rethinking the inception architecture for computer vision,'' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 2818--2826

    2016

  13. [21]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Doll \'a r et al., ``Focal loss for dense object detection,'' in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980--2988

    2017

  14. [22]

    Zhang, M

    H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, ``mixup: Beyond empirical risk minimization,'' in International Conference on Learning Representations (ICLR), 2018

    2018

  15. [23]

    S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo, and J. Choe, ``Cutmix: Regularization strategy to train strong classifiers with localizable features,'' in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2019, pp. 6023--6032

    2019

  16. [24]

    M. N. Rizve, K. Duarte, Y. S. Rawat, and M. Shah, ``In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning,'' arXiv preprint arXiv:2101.06329, 2021

    2021 arXiv

  17. [25]

    K. Wang, Y. Nie, C. Fang, C. Han, X. Wu, X. Wang, L. Lin, F. Zhou, and G. Li, ``Double-check soft teacher for semi-supervised object detection.'' in IJCAI, 2022, pp. 1430--1436

    2022

  18. [26]

    S. Li, W. Jin, Z. Wang, F. Wu, Z. Liu, C. Tan, and S. Z. Li, ``Semireward: A general reward model for semi-supervised learning,'' in ICLR, 2024

    2024

  19. [27]

    Z. Wang, Z. Chen, C. Liu, Y. Zhao, X. Zhu, and Q. J. Wu, ``Diversity augmentation and multi-fuzzy label for semi-supervised semantic segmentation,'' Neurocomputing, p. 129681, 2025

    2025

  20. [28]

    Y. Liu, Y. Tian, Y. Chen, F. Liu, V. Belagiannis, and G. Carneiro, ``Perturbed and strict mean teachers for semi-supervised semantic segmentation,'' in CVPR, 2022, pp. 4258--4267

    2022

  21. [29]

    L. Yang, L. Qi, L. Feng, W. Zhang, and Y. Shi, ``Revisiting weak-to-strong consistency in semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 7236--7246

    2023

  22. [30]

    J. Liu, Z. Jiang, T. Cao, Z. Chen, C. Zhang, and W. Gui, ``Generated pseudo-labels guided by background skeletons for overcoming under-segmentation in overlapping particle objects,'' TCSVT, vol. 33, no. 6, pp. 2906--2919, 2022

    2022

  23. [31]

    H. Hu, F. Wei, H. Hu, Q. Ye, J. Cui, and L. Wang, ``Semi-supervised semantic segmentation via adaptive equalization learning,'' NeurIPS, vol. 34, pp. 22\,106--22\,118, 2021

    2021

  24. [32]

    R. Sun, H. Mai, T. Zhang, and F. Wu, ``Daw: exploring the better weighting function for semi-supervised semantic segmentation,'' NeurIPS, vol. 36, 2024

    2024

  25. [33]

    J. Ma, C. Wang, Y. Liu, L. Lin, and G. Li, ``Enhanced soft label for semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 1185--1195

    2023

  26. [34]

    Y. Wang, H. Wang, Y. Shen, J. Fei, W. Li, G. Jin, L. Wu, R. Zhao, and X. Le, ``Semi-supervised semantic segmentation using unreliable pseudo-labels,'' in CVPR, 2022, pp. 4248--4257

    2022

  27. [35]

    B. Sun, Y. Yang, L. Zhang, M.-M. Cheng, and Q. Hou, ``Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation,'' in CVPR, 2024, pp. 3097--3107

    2024

  28. [36]

    X. Wang, M. Long, J. Wang, and M. Jordan, ``Transferable calibration with lower bias and variance in domain adaptation,'' Advances in Neural Information Processing Systems, vol. 33, pp. 19\,212--19\,223, 2020

    2020

  29. [37]

    Hoyer, D

    L. Hoyer, D. Dai, H. Wang, and L. Van Gool, ``Mic: Masked image consistency for context-enhanced domain adaptation,'' in CVPR, 2023, pp. 11\,721--11\,732

    2023

  30. [38]

    Thulasidasan, G

    S. Thulasidasan, G. Chennupati, J. A. Bilmes, T. Bhattacharya, and S. Michalak, ``On mixup training: Improved calibration and predictive uncertainty for deep neural networks,'' Advances in neural information processing systems, vol. 32, 2019

    2019

  31. [39]

    Verma, K

    V. Verma, K. Kawaguchi, A. Lamb, J. Kannala, A. Solin, Y. Bengio, and D. Lopez-Paz, ``Interpolation consistency training for semi-supervised learning,'' Neural Networks, vol. 145, pp. 90--106, 2022

    2022

  32. [40]

    Olsson, A

    C. Olsson, A. P. Eriksson, and F. Kahl, ``Improved spectral relaxation methods for binary quadratic optimization problems,'' Computer Vision and Image Understanding, vol. 112, no. 1, pp. 3--13, 2008

    2008

  33. [41]

    Tian, ``Generalized kkm theorems, minimax inequalities, and their applications,'' Journal of Optimization Theory and Applications, vol

    G. Tian, ``Generalized kkm theorems, minimax inequalities, and their applications,'' Journal of Optimization Theory and Applications, vol. 83, pp. 375--389, 1994

    1994

  34. [42]

    Everingham, S

    M. Everingham, S. A. Eslami, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman, ``The pascal visual object classes challenge: A retrospective,'' IJCV, vol. 111, pp. 98--136, 2015

    2015

  35. [43]

    Cordts, M

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Benenson, U. Franke, S. Roth, and B. Schiele, ``The cityscapes dataset for semantic urban scene understanding,'' in CVPR, 2016, pp. 3213--3223

    2016

  36. [44]

    Hariharan, P

    B. Hariharan, P. Arbeláez, L. Bourdev, S. Maji, and J. Malik, ``Semantic contours from inverse detectors,'' in ICCV, 2011, pp. 991--998

    2011

  37. [45]

    Krizhevsky, G

    A. Krizhevsky, G. Hinton et al., ``Learning multiple layers of features from tiny images,'' 2009

    2009

  38. [46]

    Vinyals, C

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra et al., ``Matching networks for one shot learning,'' Advances in neural information processing systems, vol. 29, 2016

    2016

  39. [47]

    L. Yang, W. Zhuo, L. Qi, Y. Shi, and Y. Gao, ``St++: Make self-trainingwork better for semi-supervised semantic segmentation,'' in CVPR, 2022, pp. 4258--4267

    2022

  40. [48]

    X. Lai, Z. Tian, L. Jiang, S. Liu, H. Zhao, L. Wang, and J. Jia, ``Semi-supervised semantic segmentation with directional context-aware consistency,'' in CVPR, 2021, pp. 1205--1214

    2021

  41. [49]

    S. Li, Y. He, W. Zhang, W. Zhang, X. Tan, J. Han, E. Ding, and J. Wang, ``Cfcg: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision,'' in CVPR, 2023, pp. 16\,348--16\,358

    2023

  42. [50]

    Z. Wang, Z. Zhao, X. Xing, D. Xu, X. Kong, and L. Zhou, ``Conflict-based cross-view consistency for semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 19\,585--19\,595

    2023

  43. [51]

    P. Li, P. Purkait, T. Ajanthan, M. Abdolshah, R. Garg, H. Husain, C. Xu, S. Gould, W. Ouyang, and A. Van Den Hengel, ``Semi-supervised semantic segmentation under label noise via diverse learning groups,'' in CVPR, 2023, pp. 1229--1238

    2023

  44. [52]

    H. Mai, R. Sun, T. Zhang, and F. Wu, ``Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation,'' in CVPR, 2024, pp. 3391--3401

    2024

  45. [53]

    X. Wang, B. Zhang, L. Yu, and J. Xiao, ``Hunting sparsity: Density-guided contrastive learning for semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 3114--3123

    2023

  46. [54]

    H. Wang, Q. Zhang, Y. Li, and X. Li, ``Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation,'' in CVPR, 2024, pp. 3627--3636

    2024

  47. [55]

    X. Wang, H. Bai, L. Yu, Y. Zhao, and J. Xiao, ``Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation,'' in CVPR, 2024, pp. 3303--3312

    2024

  48. [56]

    Y. Jin, J. Wang, and D. Lin, ``Semi-supervised semantic segmentation via gentle teaching assistant,'' NeurIPS, vol. 35, pp. 2803--2816, 2022

    2022

  49. [57]

    Z. Zhao, L. Yang, S. Long, J. Pi, L. Zhou, and J. Wang, ``Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 11\,350--11\,359

    2023

  50. [58]

    X. Chen, Y. Yuan, G. Zeng, and J. Wang, ``Semi-supervised semantic segmentation with cross pseudo supervision,'' in CVPR, 2021, pp. 2613--2622

    2021

  51. [59]

    Hoyer, D

    L. Hoyer, D. J. Tan, M. F. Naeem, L. Van Gool, and F. Tombari, ``Semivl: semi-supervised semantic segmentation with vision-language guidance,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2024, pp. 257--275

    2024

  52. [60]

    P. Qiao, Z. Wei, Y. Wang, Z. Wang, G. Song, F. Xu, X. Ji, C. Liu, and J. Chen, ``Fuzzy positive learning for semi-supervised semantic segmentation,'' in CVPR, 2023, pp. 15\,465--15\,474

    2023

  53. [61]

    L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and H. Adam, ``Encoder-decoder with atrous separable convolution for semantic image segmentation,'' in ECCV, 2018, pp. 801--818

    2018

  54. [62]

    K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,'' in CVPR, 2016, pp. 770--778

    2016

  55. [63]

    S. Yun, D. Han, S. Chun, S. J. Oh, Y. Yoo, and J. Choe, ``Cutmix: Regularization strategy to train strong classifiers with localizable features,'' in ICCV, 2019, pp. 6022--6031

    2019

Pith tools

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