Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

PixCon: Clean-Positive Contrastive Learning for Foundation-Model Semi-Supervised Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read With foundation-model teachers, semi-supervised segmentation gains more from clean embedding-space positives than from better pseudo-label filters.

desk verdict Clean-by-construction bank is a real, simple idea with honest analysis and a compute-matched one-switch design; the ~+0.2 per-seed lift is modest and the correctness-lever story is not yet isolated. read the letter →

arxiv 2607.03068 v1 pith:KA5BIGKC submitted 2026-07-03 cs.CV cs.AIcs.LGeess.IV

Ebenezer Tarubinga This is my paper · ORCID
classification cs.CVcs.AIcs.LGeess.IV
keywords semi-supervisedsemanticsegmentationpixelcontrastivelearningfoundationmodelsDINOv2memorybankInfoNCEcleanpositivespseudo-labelfiltering
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

Semi-supervised semantic segmentation has long asked which pseudo-labels to trust. This paper argues that foundation backbones change the operating point: a DINOv2 teacher plus a strict threshold already keeps a measured 98 percent of retained pseudo-labels clean, so leftover accuracy lives in how well the embedding space groups pixels by class. PixCon answers with a per-class memory bank that stores only labeled pixels the student already classifies correctly, forcing zero contamination by construction and adding no inference-time cost. In a compute-matched one-switch protocol over a strong consistency baseline it improves every Pascal VOC 1/8 seed by about 0.2 mIoU (three-seed mean 87.90) and matches or ties on Cityscapes and ADE20K. The authors conclude that clean-positive contrast is now a robust, low-cost default for foundation-model semi-supervised segmentation.

What carries the argument

The clean-positive bank: a per-class FIFO queue that enqueues a pixel embedding only when it is ground-truth labeled and the student already classifies it correctly. This forces ρ_F=0 by construction, so the supervised-InfoNCE update contains only true-positive signal; the paper measures real contamination under modern teachers (0.018 on Pascal, 0.106 on ADE20K) rather than assuming it.

What would settle it

Run the three-way admission-rule ablation the authors themselves flag—labeled-only versus labeled-and-correct versus confidence-filtered—on the same Pascal 1/8 seed; if labeled-only already matches labeled-and-correct, the correctness lever is not the source of the gain.

Watch

Extended reading notes

Core claim

Foundation teachers make pseudo-label noise rare under a strict threshold, so remaining accuracy is an embedding-structure problem. PixCon’s bank admits only labeled pixels the student already predicts correctly, guaranteeing contamination rate ρ_F=0; a first-order analysis shows the false-positive term of the supervised-InfoNCE gradient otherwise scales as ρ_F/(1-ρ_F). In a one-switch experiment the method lifts every Pascal-1/8 seed by roughly +0.2 mIoU while matching published UniMatch V2-B numbers and tying elsewhere, at no test-time cost.

Load-bearing premise

That also requiring the student to already be correct on a labeled pixel sharpens the true-positive contrastive signal enough to explain the measured per-seed accuracy lift, even though a clean bank and a confidence bank already tie when contamination is low.

Editorial extensions

If this is right

  • Clean-positive contrast becomes a low-cost default add-on for foundation-model consistency pipelines.
  • The ρ_F=0 guarantee matters mainly as robustness when teachers weaken; under strong teachers the accuracy comes from sharper true-positive supervision.
  • Prior contrastive SSSS methods can drop their bank-specific confidence thresholds.
  • After strict filtering the residual accuracy gap is primarily an embedding-structure problem, not a pseudo-label selection problem.

Reading between the lines

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

  • The same labeled-and-correct admission rule could transfer to other dense tasks that already use foundation backbones, such as detection or depth.
  • A soft correctness weight might recover more positives without reintroducing contamination, if the hard predicate is what currently limits bank size.
  • The measured jump in contamination from Pascal (0.018) to ADE20K (0.106) marks long-tail datasets as the natural stress test for any purity claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper argues that foundation-model teachers (DINOv2) already yield near-clean pseudo-labels under a strict threshold, so remaining SSSS gains come from structuring the embedding space rather than further filtering. It introduces PixCon, a single contrastive branch over a UniMatch-V2 consistency backbone whose per-class memory bank admits only labeled pixels that the student already classifies correctly, guaranteeing contamination rate ρ_F=0 by construction. A first-order decomposition of the supervised-InfoNCE anchor gradient shows the false-positive term scales as ρ_F/(1-ρ_F); the authors measure ρ_F (0.018 on Pascal, 0.106 on ADE20K) rather than assume it. In a compute-matched one-switch protocol the method improves every Pascal VOC 1/8 seed by ~+0.2 mIoU (3-seed mean 87.90, matching the published UniMatch V2-B figure) while Cityscapes and ADE20K remain at parity; the accuracy lift is attributed to the correctness condition sharpening the true-positive signal, with the ρ_F=0 guarantee serving mainly as robustness.

Significance. If the empirical margins and the clean-by-construction guarantee hold, PixCon supplies a low-cost, inference-free default for foundation-model SSSS that removes a tunable bank threshold and certifies positive purity. The compute-matched one-switch design, multi-seed Pascal reporting with standard deviations, direct measurement of ρ_F, controlled clean-versus-confidence ablation, and explicit negative result on adaptive thresholds are all strengths that raise the bar for subsequent work. The honesty with which the authors separate the robustness guarantee from the measured accuracy gain (and flag the missing labeled-only control) is itself a contribution to the literature.

major comments (3)
  1. [Sec. 3.3 / Table 4 / Sec. 5] Sec. 3.3, Corollary 3.2 and Sec. 4.4/Table 4: the central accuracy narrative attributes the consistent per-seed ~+0.2 mIoU lift to the correctness half of the admission rule (prediction = label) sharpening g_T. Observation 3.1 only bounds the false-positive term; once ρ_F is already ~0.02 the clean-versus-confidence banks tie within 0.02 mIoU, so contamination removal cannot be the source. The remaining causal claim is therefore asserted rather than isolated. The labeled-only versus labeled-and-correct decomposition listed as unfinished work in Sec. 5 is load-bearing for the explanation the paper advances; without it the accuracy story rests on an untested mechanism.
  2. [Table 1 / Sec. 4.2] Table 1 and Sec. 4.2: the three-seed Pascal 1/8 mean improvement reaches the published UniMatch V2-B figure, yet the sign test yields p=0.125 and the authors themselves note that part of the +0.89 gap is variance reduction (σ 0.73 → 0.26). The more reliable per-seed ~+0.2 effect is real but modest; the manuscript should either supply additional seeds or temper the claim that the three-seed mean “reaches” the published number when the comparison is not compute-matched to the full-budget published runs.
  3. [Sec. 4.2 / Tables 2–3 / Sec. 5] Sec. 4.2 and Tables 2–3: Cityscapes and ADE20K results (and several Pascal cells) are single-seed, and several splits were never run. While the paper correctly reports parity rather than over-claiming, the robustness story that “ρ_F=0 acts chiefly as teachers weaken” is left without a high-ρ_F probe (the authors flag this experiment in Sec. 5). A single controlled high-contamination run would make the gradient analysis falsifiable rather than merely consistent with the low-ρ_F regime.
minor comments (4)
  1. [Fig. 4 / Sec. 4.2] Fig. 4 caption and Sec. 4.2: the “all-live epochs” window is useful, but the early-stopping note should be moved into the main text so readers immediately understand why later epochs average over fewer seeds.
  2. [Sec. 3.3] Eq. (5)–(6) and the surrounding paragraph: assumption (iii) (comparable softmax weights) is acknowledged as an upper-bound estimate, yet the main text still presents ρ_F/(1-ρ_F) as the operative scaling. A single clarifying sentence that the ratio is an upper estimate under the stated assumptions would prevent over-reading.
  3. [Supp. H.3] Supplementary Table 10: the hyper-parameter sweeps are batch-4 / single-seed; the main text should note that the ~1 mIoU spreads are comparable to seed noise so that readers do not treat the defaults as finely optimised.
  4. [Fig. 5] Fig. 5 / Fig. 9: the red-contour visualisation is effective, but the selection criterion (“largest per-image error difference”) should be stated once in the main caption so the qualitative panel is not mistaken for a random sample.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical one-switch method paper whose gradient observation is a first-order directional argument under stated assumptions, not a fit or self-definitional prediction.

full rationale

The paper's load-bearing claims are (i) an empirical one-switch mIoU lift of PixCon over a compute-matched UniMatch V2 reimplementation (λ_pix=0) on Pascal 1/8 and (ii) a first-order decomposition of the supervised-InfoNCE anchor gradient (Obs. 3.1) showing the false-positive term scales as ρ_F/(1-ρ_F), with ρ_F measured rather than assumed. The clean-positive bank is defined by the conjunction of labeled and student-correct pixels, so ρ_F=0 holds by construction (Cor. 3.2); this is an explicit design choice, not a circular derivation of a 'prediction'. Hyperparameters (λ_pix=0.1, η=0.1, N=256) are fixed defaults with a sensitivity table; they are not fitted to force the headline number. The controlled clean-vs-confidence ablation (Table 4) is reported honestly as a near-tie, and the authors themselves flag the labeled-only vs. labeled-and-correct decomposition as unfinished future work rather than claiming it is already isolated. Self-citations (e.g., CW-BASS) appear only in related-work lists and are not load-bearing for the main claim or uniqueness. No uniqueness theorem is imported, no ansatz is smuggled via citation, and no known empirical pattern is merely renamed. The derivation chain is therefore self-contained against external benchmarks and does not reduce any claimed prediction to its own inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central empirical claim rests on standard SSSS training assumptions plus a small set of fixed contrastive hyperparameters and the modeling assumptions of Observation 3.1. No new physical entities are postulated. The free parameters are ordinary ML knobs held fixed across datasets; the load-bearing modeling choices are the InfoNCE temperature/comparability assumptions and the premise that student-correct labeled pixels sharpen g_T enough to move mIoU.

free parameters (6)
  • λ_pix (contrastive loss weight) = 0.1
    Fixed at 0.1 for all datasets; sensitivity table shows off-default values hurt. Chosen by hand, not derived.
  • η (InfoNCE temperature) = 0.1
    Fixed at 0.1; standard contrastive hyperparameter, not derived from data.
  • N (per-class bank size) = 256
    Fixed at 256 entries per class; sensitivity shows nearby values slightly worse.
  • m (anchors per class cap) and max anchors per iter = m=64, max=1024
    m=64, max 1024; compute caps chosen for training efficiency.
  • τ (consistency confidence threshold) = 0.95
    Inherited from UniMatch V2 at 0.95; not re-tuned for PixCon but load-bearing for the measured ρ_F and the baseline.
  • D (projection dimension) = 256
    Embedding dim 256 for the 2×1×1 head; standard design choice.
assumptions (5)
  • ad hoc to paper ℓ2-normalized embeddings and shared temperature make per-entry softmax weights of true and false positives comparable (Obs. 3.1 assumption iii).
    Load-bearing for the ρ_F/(1−ρ_F) scaling; the paper itself notes it likely overstates g_F in a well-clustered regime.
  • domain assumption DINOv2-B with a strict global threshold already retains a mostly-clean pseudo-label set, so remaining accuracy lives in embedding structure rather than filter design.
    Stated in abstract and Introduction; supported by measured ρ_F≈0.018 on Pascal but is the regime premise for the whole paper.
  • domain assumption Supervised InfoNCE / SupCon is a valid pixel-level objective for structuring decoder embeddings in SSSS.
    Inherited from Alonso et al., ReCo, U²PL; not re-derived.
  • domain assumption EMA teacher with weak-to-strong consistency (UniMatch V2) is a valid substrate; switching λ_pix=0 recovers the baseline.
    Standard SSSS practice; used as the one-switch control.
  • standard math Standard optimization and data-augmentation facts (AdamW, poly LR, CutMix, channel dropout) do not interact pathologically with the clean bank.
    Background training assumptions shared with UniMatch V2.
invented entities (2)
  • Clean-positive memory bank (admission: labeled ∧ student-correct)
    purpose: Guarantee ρ_F=0 positives for supervised InfoNCE and sharpen true-positive signal without a bank threshold.
    Core methodological object. Independent evidence is the ablation and multi-seed mIoU; no external physical prediction.
  • Contamination rate ρ_F as fraction of false positives in the positive set independent evidence
    purpose: Quantify how much a confidence bank pollutes the InfoNCE gradient; measured on val under τ=0.95.
    Definitional quantity, not a new physical entity; useful bookkeeping for the gradient argument.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PixCon: Clean-Positive Contrastive Learning for Foundation-Model Semi-Supervised Segmentation." pith.science (2026). https://pith.science/paper/KA5BIGKC

@misc{pith2026260703068,
  author       = {Pith},
  title        = {Pith review of: PixCon: Clean-Positive Contrastive Learning for Foundation-Model Semi-Supervised Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KA5BIGKC}},
  note         = {Machine review of arXiv:2607.03068}
}
abstract

Semi-supervised semantic segmentation (SSSS) has long turned on one question, which pseudo-labels to trust, and answered it with ever more careful confidence filtering. Foundation backbones change the regime: with a DINOv2 teacher a strict threshold already retains a measured 98%-clean pseudo-label set, so the accuracy that remains lives not in the filter but in how the embedding space is structured by class. We propose PixCon, a clean-positive pixel-contrastive framework. PixCon maintains a per-class memory bank that admits only labeled pixels the student already classifies correctly, guaranteeing a contamination-free positive set ($\rho_F=0$) by construction, unlike prior contrastive SSSS banks (ReCo, U$^2$PL) built from confidence-filtered pseudo-labels. It is a single branch over a consistency backbone, adds no inference-time parameters, and needs no bank-specific threshold. A first-order analysis of the supervised-InfoNCE gradient explains why contamination hurts: its false-positive term scales as $\rho_F/(1-\rho_F)$, which we measure (0.018 on Pascal, 0.106 on ADE20K) rather than assume. Across Pascal VOC, Cityscapes, and ADE20K, PixCon matches or improves a strong DINOv2-based UniMatch V2 baseline in a compute-matched one-switch protocol: it improves every Pascal-1/8 seed (a per-seed gain of about +0.2 mIoU) and its three-seed mean reaches 87.90, the published UniMatch V2-B figure. Because contamination is already rare under foundation-model teachers, our analysis indicates the $\rho_F=0$ guarantee acts chiefly as robustness as teachers weaken, while the accuracy gain comes from cleaner positive supervision, making clean-positive contrast a robust, low-cost default for foundation-model SSSS.

Figures

Figures reproduced from arXiv: 2607.03068 by the authors.

Figure 1
Figure 1. One switch, cleaner supervision. A selected ADE20K 1/8 win (input | GT | UniMatch V2 | PixCon; red = error vs. GT). PixCon’s clean-positive ρF=0 branch, one switch over Uni￾Match V2, leaves far fewer error contours while matching it in aggregate (Pascal 1/8 87.90, a 3-seed mean) at no test-time cost. Full set [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The PixCon architecture. PixCon couples two branches over a shared DINOv2-B encoder and DPT decoder, trained end-to￾end under one objective. Consistency branch (top). A weak view of an unlabeled image passes through the EMA teacher to produce a pseudo-label yˆ and confidence c; a strict mask M: c≥0.95 filters it, and two CutMix strong views with complementary channel dropout are trained to agree (Lu). This branch ad… view at source ↗
Figure 3
Figure 3. Why clean positives matter. Schematic of the supervised-InfoNCE gradient on an anchor za of class a. (a) A confidence-filtered bank (ReCo, U2 PL) admits a small fraction ρF of confidently-wrong entries, pixels of another class b enqueued under a. These act as false positives and contribute a gradient component (red) pulling za toward the wrong-class region, so the net update (black) is deflected. (b) PixCon admits a… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Consistent per-epoch margin (3 seeds). EMA-teacher mIoU over training (Pascal, DINOv2-Base; thin lines per seed, bold the 3-seed mean). At 1/8 (right) the PixCon (green) band sits above UniMatch V2 (blue) across seeds and leads in 41/41 all￾live epochs; at 1/16 (left) …
Figure 5
Figure 5. Figure 5: Qualitative wins on Pascal VOC 1/8. Six per-image PixCon-vs-UniMatch V2 comparisons at the cell of our robust multi-seed gain (DINOv2-Base, EMA-teacher, seed 0), as blocks of input | GT | UniMatch V2 repro | PixCon; red contours outline disagreement with GT (void ignor…
Figure 6
Figure 6. Figure 6: ADE20K 1/8: eight PixCon wins. Per-image PixCon-vs-UniMatch V2 comparisons (150-class palette), ranked by per-image error difference, in the same input | GT | UniMatch V2 | PixCon format as [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: The bottleneck has moved to the encoder, and Pix￾Con adds a lever on top of it. Pascal VOC 1/8 mIoU versus total model size (log axis) for representative SSSS methods, grouped by backbone family (prior-work accuracies from the published num￾bers; prior-work parameter c…
Figure 8
Figure 8. Figure 8: PixCon training step (one iteration). The clean-positive filter (valid) keeps only labeled pixels the student already classifies correctly; the bank is enqueued after the loss so anchors never appear in their own positive set. E. Full Per-Class Results The per-class Io…
Figure 9
Figure 9. Figure 9: Complete qualitative set: sixteen PixCon wins. All sixteen Pascal VOC 1/8 validation images with the largest per-image error advantage of PixCon over the UniMatch V2 reproduction (DINOv2-Base, EMA-teacher predictions, seed 0), as two side-by-side blocks of input | grou…
Figure 10
Figure 10. Figure 10: Cityscapes qualitative (1/16): near-identical, PixCon marginally cleaner. Six Cityscapes val images with the largest per-image error advantage of PixCon over the UniMatch V2 reproduction (DINOv2-Base, EMA-teacher predictions, seed 0), as two side￾by-side blocks of inp…
Figure 11
Figure 11. Figure 11: Feature geometry of the shared fused decoder embedding (Pascal 1/8; seed 0 shown, metrics in [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: The gain is broad, not rare-class-specific. Per-class IoU change (PixCon − UniMatch V2 reproduction), mean over 3 seeds with ±std error bars, on Pascal VOC 1/16 (left) and 1/8 (right), DINOv2-Base, classes sorted by baseline difficulty (hardest at top). Green bars are…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A saturation gate measures the reliability of a teacher's confident pseudo-labels, Pr(correct | confidence >= 0.95), and picks strict filtering when that reliability is high and an adaptive floor when it drops, making...

Reference graph

Works this paper leans on

38 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Iñigo Alonso, Alberto Sabater, David Ferstl, Luis Monte- sano, and Ana C. Murillo. Semi-supervised semantic seg- mentation with pixel-level contrastive learning from a class- wise memory bank. InICCV, 2021. 2

    2021

  2. [2]

    Learning imbalanced datasets with label- distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label- distribution-aware margin loss. InNeurIPS, 2019. 3

    2019

  3. [3]

    Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning

    Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi- supervised learning. InICLR, 2023. 3

    2023

  4. [4]

    Semi-supervised semantic segmentation needs strong, varied perturbations

    Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation needs strong, varied perturbations. InCVPR, 2021. 1, 2

    2021

  5. [5]

    Feedback-driven pseudo- label reliability assessment: Redefining thresholding for semi-supervised semantic segmentation.arXiv preprint arXiv:2505.07691, 2025

    Negin Ghamsarian, Sahar Nasirihaghighi, Klaus Schoeff- mann, and Raphael Sznitman. Feedback-driven pseudo- label reliability assessment: Redefining thresholding for semi-supervised semantic segmentation.arXiv preprint arXiv:2505.07691, 2025. 3

    arXiv 2025

  6. [6]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InCVPR,

  7. [7]

    Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation

    Rui He, Jie Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. InCVPR, 2022. 3, 5

    2022

  8. [8]

    Beyond pixels: Semi-supervised semantic segmenta- tion with a multi-scale patch-based multi-label classifier

    Prantik Howlader, Srijan Das, Hieu Le, and Dimitris Sama- ras. Beyond pixels: Semi-supervised semantic segmenta- tion with a multi-scale patch-based multi-label classifier. In ECCV, 2024. 2, 7, 8

    2024

Show all 38 references
  1. [9]

    SemiVL: Semi-supervised semantic segmenta- tion with vision-language guidance

    Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. SemiVL: Semi-supervised semantic segmenta- tion with vision-language guidance. InECCV, 2024. 2, 7, 8

    2024

  2. [10]

    Semi-supervised semantic segmentation via adaptive equalization learning

    Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. InNeurIPS, 2021. 3

    2021

  3. [11]

    Semi-supervised semantic segmentation via gentle teaching assistant

    Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. In NeurIPS, 2022. 2

    2022

  4. [12]

    CAFS: Class adaptive framework for semi-supervised semantic segmentation.arXiv preprint arXiv:2303.11606, 2023

    Jingi Ju, Hyeoncheol Noh, Yooseung Wang, Minseok Seo, and Dong-Geol Choi. CAFS: Class adaptive framework for semi-supervised semantic segmentation.arXiv preprint arXiv:2303.11606, 2023. 3

    2023 arXiv

  5. [13]

    Supervised contrastive learning

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. InNeurIPS,

  6. [14]

    Ro- bust pseudo-label learning for semantic segmentation: An encoding perspective.arXiv preprint arXiv:2512.06870,

    Wangkai Li, Rui Sun, Zhaoyang Li, and Tianzhu Zhang. Ro- bust pseudo-label learning for semantic segmentation: An encoding perspective.arXiv preprint arXiv:2512.06870,

  7. [15]

    Shikun Liu, Shuaifeng Zhi, Edward Johns, and Andrew J. Davison. Bootstrapping semantic segmentation with re- gional contrast. InICLR, 2022. 2

    2022

  8. [16]

    Perturbed and strict mean teachers for semi-supervised semantic segmenta- tion

    Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmenta- tion. InCVPR, 2022. 2, 7

    2022

  9. [17]

    Improving semi-supervised semantic segmentation with sliced-wasserstein feature alignment and uniformity

    Chen-Yi Lu, Kasra Derakhshandeh, and Somali Chaterji. Improving semi-supervised semantic segmentation with sliced-wasserstein feature alignment and uniformity. In CVPR, 2025. 3

    2025

  10. [18]

    RankMatch: Exploring the better consistency regularization for semi-supervised semantic segmentation

    Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. RankMatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. InCVPR, 2024. 2

    2024

  11. [19]

    V o, Marc Szafraniec, et al

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V . V o, Marc Szafraniec, et al. Dinov2: Learning robust visual features without supervision.TMLR, 2024. 1, 2, 5

    2024

  12. [20]

    Vi- sion transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InICCV, 2021. 2, 5

    2021

  13. [21]

    PrevMatch: Revisiting and maximizing temporal knowledge in semi-supervised semantic segmenta- tion

    Wooseok Shin et al. PrevMatch: Revisiting and maximizing temporal knowledge in semi-supervised semantic segmenta- tion. InWACV, 2026. 2, 7

    2026

  14. [22]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ankit Singh Rawat, Omry Shavit, et al. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. InNeurIPS, 2020. 1, 3

    2020

  15. [23]

    DAW: Exploring the better weighting function for semi-supervised semantic segmentation

    Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. DAW: Exploring the better weighting function for semi-supervised semantic segmentation. InNeurIPS, 2023. 2

    2023

  16. [24]

    CorrMatch: Label propagation via correlation matching for semi-supervised se- mantic segmentation

    Zesen Sun, Fangrui Yang, Qi Hu, et al. CorrMatch: Label propagation via correlation matching for semi-supervised se- mantic segmentation. InCVPR, 2024. 2, 6, 7, 8

    2024

  17. [25]

    CW-BASS: Confidence-weighted boundary-aware learning for semi-supervised semantic segmentation

    Ebenezer Tarubinga, Jenifer Kalafatovich, and Seong-Whan Lee. CW-BASS: Confidence-weighted boundary-aware learning for semi-supervised semantic segmentation. In 2025 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2025. 2, 7 9

    2025

  18. [26]

    Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. InNeurIPS, 2017. 2

    2017

  19. [27]

    Seesaw loss for long- tailed instance segmentation

    Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yizhuo Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long- tailed instance segmentation. InCVPR, 2021. 3

    2021

  20. [28]

    Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation

    Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In CVPR, 2024. 2, 7

    2024

  21. [29]

    Semi-supervised semantic segmentation using unreliable pseudo-labels

    Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. InCVPR, 2022. 2, 6, 7

    2022

  22. [30]

    Freematch: Self-adaptive thresholding for semi-supervised learning

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. InICLR, 2023. 3

    2023

  23. [31]

    CReST: A class-rebalancing self-training frame- work for imbalanced semi-supervised learning

    Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. CReST: A class-rebalancing self-training frame- work for imbalanced semi-supervised learning. InCVPR,

  24. [32]

    Allspark: Reborn labeled features from unlabeled in trans- former for semi-supervised semantic segmentation

    Haonan Yang, Muxin Li, Yunzhe Zhuge, and Hao Lu. Allspark: Reborn labeled features from unlabeled in trans- former for semi-supervised semantic segmentation. In CVPR, 2024. 2, 6, 7

    2024

  25. [33]

    St++: Make self-training work better for semi-supervised se- mantic segmentation

    Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised se- mantic segmentation. InCVPR, 2022. 1, 2, 6, 7, 12

    2022

  26. [34]

    Revisiting weak-to-strong consistency in semi-supervised semantic segmentation

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. InCVPR, 2023. 2, 3, 6, 7, 8

    2023

  27. [35]

    UniMatch V2: Pushing the limit of semi- supervised semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. UniMatch V2: Pushing the limit of semi- supervised semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025. 1, 2, 3, 4, 5, 6, 7, 8

    2025

  28. [36]

    Flexmatch: Boosting semi-supervised learning with curricu- lum pseudo labeling

    Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jin- dong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curricu- lum pseudo labeling. InNeurIPS, 2021. 3

    2021

  29. [37]

    Instance-specific and model-adaptive supervision for semi-supervised semantic segmentation

    Zhen Zhao, Sifan Long, Jimin Pi, Jingdong Wang, and Lup- ing Zhou. Instance-specific and model-adaptive supervision for semi-supervised semantic segmentation. InCVPR, 2023. 2

    2023

  30. [38]

    rare classes have lower confidence and should be given lower thresholds to admit more of their pseudo-labels

    Ziming Zhao, Shuai Yang, Haoyu Xing, Siyu Xu, Yupeng Yang, and Yao Zhang. Augseg: Maximizing the utility of unlabeled data for semi-supervised semantic segmentation. InCVPR, 2023. 2, 6, 7, 8 10 Supplementary Material A. Extended Gradient-Quality Analysis This appendix expands ...

    2023

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.