Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation

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

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

Pith's one-line read The paper claims that Shuffle PatchMix augmentation and confidence-margin pseudo-label weighting achieve new state-of-the-art results for source-free domain adaptation on PACS, VisDA-C, and DomainNet-126.

desk verdict Useful but incremental SFDA paper; the headline PACS gain is suspect because the AdaContrast baseline likely ran for half the epochs of the proposed method. read the letter →

arxiv 2505.24216 v1 pith:WZJCWPI6 submitted 2025-05-30 cs.CV

Prasanna Reddy Pulakurthi , Majid Rabbani , Jamison Heard , Sohail Dianat , Celso M. de Melo , Raghuveer Rao This is my paper
classification cs.CV
keywords source-freedomainadaptationpseudo-labelreweightingdataaugmentationShufflePatchMixcontrastivelearningself-trainingshiftimageclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to show that source-free domain adaptation—adapting a model to a new unlabeled target domain without access to the original source data—can be made substantially more accurate by changing two things in a standard self-training pipeline. First, it replaces plain strong augmentation with Shuffle PatchMix, which shuffles and blends patches within each target image to create harder, more diverse views. Second, it weights each pseudo-label by a confidence-margin score so that uncertain labels contribute less to training. On three benchmarks the combined method reports the highest average accuracies: 86.7% on PACS (7.3 points above a reproduced baseline), 89.4% on VisDA-C, and 71.1% on DomainNet-126. If these results hold, the practical payoff is that a deployed model can adapt to a new domain using only unlabeled data and cheap augmentations.

What carries the argument

The engine of the method is a modified AdaContrast-style self-training loop with two inserted components. Shuffle PatchMix (SPM) cuts each target image into $\nu$ random patches, shuffles them, and linearly blends the shuffled version with the original using $\lambda\sim\mathrm{Beta}(a,b)$, then applies standard strong augmentations; overlapping patches (extracted 30% larger than the intended size) and a schedule that decreases $a$ during training make the mixes smooth and progressively stronger. The second component is a continuous pseudo-label weight $w_{x_t}=p_{\text{top1}}\,\Delta\,e^{\Delta}$ computed from the confidence and margin of nearest-neighbor-refined pseudo-labels, used to scale the classification loss; the contrastive and diversity losses from the base framework are kept. The weight formula is what carries the noise-reduction claim, and SPM is what carries the diversity claim.

What would settle it

Take a target set with known labels, compute the trustworthiness score for each image's guessed label, and check whether the lowest-scoring quarter of images actually has the lowest guessed-label accuracy; if not, the reweighting is not ordering samples by reliability.

Watch

Extended reading notes

Core claim

The central discovery is that intra-image patch mixing and pseudo-label reweighting are complementary fixes for the two failure modes of pseudo-label self-training: overfitting on small target sets and amplification of label noise. The paper argues that SPM's overlapping, shuffled patch blends, with a decreasing Beta mixing strength over training, generate augmentations that are challenging but still realistic, pushing the encoder to learn class-invariant features. The weighting term $w_{x_t}=p_{\text{top1}}\,\Delta\,e^{\Delta}$, where $\Delta=p_{\text{top1}}-p_{\text{top2}}$ on nearest-neighbor-refined pseudo-labels, continuously down-weights uncertain samples. Under this recipe the method improves the baseline by 7.3 points on PACS single-target, 7.2 points on PACS multi-target, 2.8 points on DomainNet-126, and 0.7 points on VisDA-C, setting the reported best average accuracy on all three benchmarks.

Load-bearing premise

The method's bet is that its trustworthiness score for an automatically guessed label—top-1 probability times top-1/top-2 gap times the exponential of that gap—correctly identifies which guesses are reliable, so that down-weighting the low scorers removes noise rather than useful information.

Editorial extensions

If this is right

  • If the central claim is right, source-free adaptation on small target datasets becomes more practical, since gains of more than 7 points on PACS suggest overfitting, not just domain shift, was the binding constraint.
  • A single unlabeled target stream can be adapted without source data or target labels, using only intra-image augmentations and confidence-weighted pseudo-labels.
  • The recipe keeps the same contrastive and diversity losses, so it can be dropped into existing AdaContrast-style SFDA pipelines by swapping the strong augmentation and scaling the classification loss.
  • The continuous weighting removes the need for hand-tuned confidence thresholds, which prior methods had to set explicitly.

Reading between the lines

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

  • The paper leaves the exponential factor in $w_{x_t}$ unablated; a natural next test is whether a simpler monotone weight, such as $p_{\text{top1}}\Delta^2$, gives the same gains, which would shift the credit from the exact formula to the general idea of margin-aware weighting.
  • Because SPM is intra-image and needs no source-target pairs, the augmentation could plausibly transfer to semi-supervised or self-supervised learning, though the paper only evaluates source-free domain adaptation.
  • The largest gains appear on the smallest dataset (PACS), which implies a testable prediction: the benefit of SPM should shrink as target-data size grows, and a reader could check this by subsampling VisDA-C or DomainNet.
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 presents two modifications to the AdaContrast source-free domain adaptation (SFDA) framework: (1) Shuffle PatchMix (SPM), an intra-image augmentation that randomly shuffles patches, blends them via a Beta-distributed mixing coefficient, and uses overlapping patch blending to reduce blocking artifacts, and (2) a Confidence-Margin reweighting of pseudo-labels, where each sample's classification loss is weighted by w = p_top1 * Δ * exp(Δ), with Δ the margin between top-1 and top-2 pseudo-label probabilities. The method is evaluated on PACS, VisDA-C, and DomainNet-126, reporting state-of-the-art averages of 86.7%, 89.4%, and 71.1% respectively, with ablations on PACS and DomainNet-126. The central claim is that SPM and the reweighting strategy jointly drive large gains, particularly on small datasets such as PACS.

Significance. If the empirical results hold, the paper would provide a simple, practical improvement over AdaContrast with public code, and the PACS gains would be notable. The authors are to be credited for releasing code, reproducing the AdaContrast baseline, and providing component ablations. However, the evidence as presented is not yet at the standard needed for the claimed new benchmarks: the main PACS comparison may be confounded by an unmatched training budget, the adaptive mixing schedule is under-specified and justified only by an unseen simulation, the weighting function is not ablated against alternatives, and all results are single runs with no error bars. These issues are fixable within the scope of the paper, but they currently leave the headline claims under-supported.

major comments (4)
  1. [§4 (Hyperparameters), Table 1, Table 5] The central PACS comparison is potentially confounded by unequal training budgets. Section 4 states that the proposed model is trained for 100 epochs on PACS, while the AdaContrast baseline is described only as 'reproduced using [6]' with no reported epoch count. If the reproduced baseline follows the default AdaContrast protocol of 50 epochs, the +7.3% in Table 1 and the incremental gains in the PACS ablation rows of Table 5 combine the effect of the proposed components with that of doubling the training length. Please rerun AdaContrast under an identical 100-epoch (or otherwise matched) budget and report the epoch count for every baseline and ablation configuration.
  2. [§3 (Adaptive Mixing Strength) and §4 (Hyperparameters)] The adaptive mixing schedule is a core component of SPM, yet it is not reproducible from the manuscript. Only the start value a_s is reported; the end value a_e, the decay shape, and b are missing. The only justification for decreasing the Beta parameter is the sentence 'Our simulations demonstrate...' with no figure, table, or appendix. Please specify a_e and b, describe how a decreases over training, and provide the simulation evidence (or remove the claim if it is not needed).
  3. [§3 (Eq. 2) and Table 5] The reweighting function w = p_top1 Δ exp(Δ) is not validated against alternative forms. Table 5 switches the reweighting module on and off but does not compare, for example, w = p_top1, w = p_top1 Δ, or a confidence-threshold filter such as UPA's. Thus the exponential factor, which the text justifies as making 'statistically the most trustworthy' samples dominate, is an untested design choice. Please add ablations over the functional form of the weight and, ideally, a direct analysis of how w correlates with pseudo-label correctness.
  4. [Tables 1-5] All quantitative results appear to be single runs without error bars or multiple seeds. Because the method has stochastic components (Beta sampling, random patch shuffle, nearest-neighbor queue) and the hyperparameters are tuned per dataset (Section 4), a single run does not establish that the reported margins—0.7% on VisDA-C, 2.8% on DomainNet-126, and even the larger PACS gains—are significant. Please report mean and standard deviation over at least three seeds, and include sensitivity analyses for ρ and the other reported hyperparameters rather than only the tuned configuration.
minor comments (5)
  1. [§3 and §7] The supplement link appears as a placeholder '(link)' and should be a real URL.
  2. [Throughout] The dataset name is inconsistent: 'VISDA-C' appears in Section 2 and the Table 3 caption, while 'VisDA-C' is used elsewhere; please unify the spelling.
  3. [Table 5] Table 5 is presented as a prose list rather than a conventional table; please format it with explicit rows and checkmarks so the configurations are unambiguous.
  4. [§4] The phrase 'Since [6] had not originally reported results for PACS' is awkward; suggest 'Because AdaContrast [6] did not report PACS results...'.
  5. [Table 3] The caption promises underlining for the second-highest values, but no values appear underlined in the table; either format them accordingly or adjust the caption.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the SOTA claims are empirical comparisons against external benchmarks; the PACS baseline training-budget ambiguity and the unpublished 'simulations' claim are correctness/support gaps, not derivation-to-input reductions.

full rationale

The paper makes no claim that derives a target result from a defining formula. The two contributions — Shuffle PatchMix and the Confidence-Margin weighting w_xt = p_top1 · Δ · exp(Δ) (Eq. 2) — are heuristic design choices evaluated empirically, and the central 'state-of-the-art' claim is tested against external published numbers for VisDA-C and DomainNet-126 and against a reproduced AdaContrast baseline for PACS. No quantity fitted to data is later renamed as a prediction: the ablation table reports actual runs, and the +7.3% PACS headline is a measured difference, not a construction. Two concerns therefore fall outside the circularity definition. First, the PACS baseline ('the publicly available version of the code (after being verified on the published results in [6]) was used to generate the baseline PACS results', Section 4) is reproduced at an unstated epoch count while the proposed method is 'trained for 100 epochs for PACS', so the headline gain may partially reflect an unmatched training budget; this is a benchmarking-fairness risk, not a circular reduction, and the method still exceeds the external NEL result on PACS by 14.3%. Second, the adaptive mixing schedule is justified only by 'Our simulations demonstrate that gradually decreasing the mean of the random variable λ during the training ... allows the model to better adapt to the SPM augmentation distribution' — a missing-support assertion with no shown evidence, and the exponential term in Eq. 2 is not ablated against alternative weighting functions; these are support gaps, not self-definitional steps. The only self-citation (ref. [5]) appears in a background list for UDA feature alignment and carries no load. Verdict: no significant circularity.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The method relies on standard SFDA assumptions plus several hand-chosen design elements, including the Beta schedule, rho, patch counts, and an unvalidated weighting function. No new physical or conceptual entities are postulated.

free parameters (9)
  • Beta distribution scale parameter a_s = 8 for DomainNet-126/PACS, 4 for VisDA-C
    Controls mixing strength; chosen per dataset to maximize reported accuracy.
  • Beta distribution parameter b = not explicitly reported (figure shows b=2)
    Required to reproduce the augmentation; absent from hyperparameters.
  • a_e (end value for adaptive mixing schedule) = not reported
    Adaptive mixing is defined as decreasing a from a_s to a_e, but a_e is never given.
  • rho (fraction of SPM augmentations) = 0.8
    Chosen to balance diversity and realism; no sensitivity analysis shown.
  • patch count nu = sampled from {4, 16, 64, 256}
    Random patch grid sizes; not justified beyond diversity.
  • number of nearest neighbors k = 3
    Used for pseudo-label refinement; deviates from AdaContrast default.
  • learning rate = 2e-4
    Tuned for optimal performance.
  • training epochs = 100 for PACS, 50 for DomainNet-126 and VisDA-C
    Extended training claimed necessary for adaptive mixing schedule.
  • exponential margin factor exp(delta) = implicit in Equation 2
    Ad hoc gain function; no derivation or ablation on alternative weighting functions.
assumptions (5)
  • standard math Beta distribution formula for mixing parameter lambda
    Used in Equation 1 to generate the mixing strength.
  • domain assumption Source model provides a useful starting point for target pseudo-labels
    Standard SFDA assumption; entire self-training relies on initial pseudo-label quality.
  • domain assumption Nearest-neighbor soft voting in a memory queue produces better pseudo-labels than direct predictions
    Adopted from AdaContrast [6]; no analysis in this paper.
  • ad hoc to paper Decreasing the mean of lambda during training helps the model adapt to SPM
    Asserted from unpublished simulations; no evidence included in the text.
  • ad hoc to paper Confidence-margin product p_top1 * margin * exp(margin) orders samples by reliability
    Heuristic introduced without derivation or independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation." pith.science (2026). https://pith.science/paper/WZJCWPI6

@misc{pith2026250524216,
  author       = {Pith},
  title        = {Pith review of: Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZJCWPI6}},
  note         = {Machine review of arXiv:2505.24216}
}
read the original abstract

This work investigates Source-Free Domain Adaptation (SFDA), where a model adapts to a target domain without access to source data. A new augmentation technique, Shuffle PatchMix (SPM), and a novel reweighting strategy are introduced to enhance performance. SPM shuffles and blends image patches to generate diverse and challenging augmentations, while the reweighting strategy prioritizes reliable pseudo-labels to mitigate label noise. These techniques are particularly effective on smaller datasets like PACS, where overfitting and pseudo-label noise pose greater risks. State-of-the-art results are achieved on three major benchmarks: PACS, VisDA-C, and DomainNet-126. Notably, on PACS, improvements of 7.3% (79.4% to 86.7%) and 7.2% are observed in single-target and multi-target settings, respectively, while gains of 2.8% and 0.7% are attained on DomainNet-126 and VisDA-C. This combination of advanced augmentation and robust pseudo-label reweighting establishes a new benchmark for SFDA. The code is available at: https://github.com/PrasannaPulakurthi/SPM

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 47 canonical work pages

  1. [6]

    AugmentationOverlap Strategy Avg. Avg. ✓× × × 67.8 79.4 ✓× ×✓ 69.1 81.8 ✓ ✓× × 70.2 83.8 ✓ ✓ ✓× 70.4 84.4 ✓ ✓ ✓ ✓ 71.1 86.7 are combined, the highest accuracy of71.1%on DomainNet- 126 and86.7%on PACS is achieved, demonstrating their complementary benefits in improving SFDA performance. The largest improvement is observed in the smaller dataset, PACS, due ...

    work page

  2. [1]

    However, domain shifts can degrade network perfor- mance considerably [1]

    INTRODUCTION Deep neural networks have achieved remarkable success in tasks where the training and test data share similar distribu- tions. However, domain shifts can degrade network perfor- mance considerably [1]. To address this challenge, Domain Adaptation (DA) techniques have been developed to improve model performance in the target domain by leveragi...

    work page

  3. [2]

    Shuffle PatchMix Augmentation with Confidence-Margin Weighted Pseudo-Labels for Enhanced Source-Free Domain Adaptation

    RELATED WORK Domain Adaptation:In UDA, several methods have been proposed to reduce domain discrepancies by aligning the feature distributions between source and target domains, MMD [7], MCC [8], MDD [9], and CMD [10]. Addition- ally, GAN-based methods use adversarial training to align the distributions in feature space [2, 4] and image space [11]. Howeve...

    work page Pith review arXiv 2025

  4. [3]

    METHOD In the SFDA setting, the source modelg s(.)is trained on la- beled source data{x i s, yi s}ns i=1, wherex i s ∈X s represent the Pseudo Labels Encoder Pseudo- Label Refinement Classifier Encoder Momentum Encoder Similarity 𝐿𝑐𝑒 𝐿𝑑𝑖𝑣 𝐿𝑐𝑡𝑟 𝒙𝒕 : Target Image 𝑡𝑠(𝒙𝒕): SPM Aug Queue Shared 𝐵𝑒𝑡𝑎(𝑎, 𝑏) 𝑡𝑤(𝒙𝒕): Weak Aug 𝑘 𝑞 {𝑘+, 𝑘1 −, 𝑘2 −, … 𝑘𝑀 − } 𝑡′𝑠(𝒙𝒕):...

    work page

  5. [4]

    EXPERIMENTAL SETUP Datasets:The PACS [33], VisDA-C [34], and DomainNet- 126 [35] datasets are used to evaluate the proposed method. For DomainNet-126, seven domain shifts were constructed from four domains (Real, Sketch, Clipart, Painting) follow- ing the protocol described in [6], with the top-1 accuracy (%) and the average of the seven shifts reported. ...

    work page

  6. [5]

    For the single-target setting, our method achieves an average accuracy of86.7%, outperforming the baseline AdaContrast [6] by a significant margin of7.3%

    RESULTS PACS Results:Table 1 and Table 2 summarize the perfor- mance of our proposed method on the PACS dataset under single-target and multi-target settings. For the single-target setting, our method achieves an average accuracy of86.7%, outperforming the baseline AdaContrast [6] by a significant margin of7.3%. The most notable improvements over base- li...

    work page

  7. [7]

    Learning transferable features with deep adaptation networks,

    Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan, “Learning transferable features with deep adaptation networks,” in ICML. PMLR, 2015, pp. 97–105

    work page 2015

  8. [8]

    Experimental results demonstrate that the proposed approach outperforms state-of-the-art methods across three benchmark datasets

    CONCLUSION This paper introduces the SPM augmentation technique and the Confidence-Margin reweighting strategy to advance SFDA. Experimental results demonstrate that the proposed approach outperforms state-of-the-art methods across three benchmark datasets. Notably, the method is especially ef- fective on smaller datasets like PACS, where limited target s...

    work page

Show all 49 references
  1. [9]

    Joaquin Quionero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil Lawrence,Dataset Shift in Machine Learning, The MIT Press, 2009

    2009

  2. [10]

    Adver- sarial discriminative domain adaptation,

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell, “Adver- sarial discriminative domain adaptation,” inCVPR, July 2017

    2017

  3. [11]

    Unsupervised domain adapta- tion by backpropagation,

    Yaroslav Ganin and Victor Lempitsky, “Unsupervised domain adapta- tion by backpropagation,” inICML. PMLR, 2015, pp. 1180–1189

    2015

  4. [12]

    Conditional adversarial domain adaptation,

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan, “Conditional adversarial domain adaptation,”NeurIPS, vol. 31, 2018

    2018

  5. [13]

    Unsupervised domain adaptation us- ing feature aligned maximum classifier discrepancy,

    Prasanna Reddy Pulakurthi, Sohail A Dianat, Majid Rabbani, Suya You, and Raghuveer M Rao, “Unsupervised domain adaptation us- ing feature aligned maximum classifier discrepancy,” inApplications of Machine Learning 2022. SPIE, 2022, vol. 12227, pp. 37–45

    2022

  6. [14]

    Con- trastive test-time adaptation,

    Dian Chen, Dequan Wang, Trevor Darrell, and Sayna Ebrahimi, “Con- trastive test-time adaptation,” inCVPR, 2022, pp. 295–305

    2022

  7. [15]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,

    Jian Liang, Dapeng Hu, and Jiashi Feng, “Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,” inICML. PMLR, 2020, pp. 6028–6039

    2020

  8. [16]

    Min- imum class confusion for versatile domain adaptation,

    Ying Jin, Ximei Wang, Mingsheng Long, and Jianmin Wang, “Min- imum class confusion for versatile domain adaptation,” inECCV. Springer, 2020, pp. 464–480

    2020

  9. [17]

    Bridging theory and algorithm for domain adaptation,

    Yuchen Zhang, Tianle Liu, Mingsheng Long, and Michael Jordan, “Bridging theory and algorithm for domain adaptation,” inICML. PMLR, 2019, pp. 7404–7413

    2019

  10. [18]

    Central moment discrep- ancy for domain-invariant representation learning,

    Werner Zellinger, Thomas Grubinger, Edwin Lughofer, Thomas Natschl¨ager, and Susanne Saminger-Platz, “Central moment discrep- ancy for domain-invariant representation learning,” inICLR, 2017

    2017

  11. [19]

    While this method im- proved the robustness of pseudo-labels, it incurred a signif- icant computational cost due to the need for multiple aug- mentations and predictions

    method employed a pseudo-label refinement framework that relied on an ensemble of pseudo-label predictions gen- erated from different augmentations. While this method im- proved the robustness of pseudo-labels, it incurred a signif- icant computational cost due to the need for...

  12. [20]

    CyCADA: Cycle- consistent adversarial domain adaptation,

    Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell, “CyCADA: Cycle- consistent adversarial domain adaptation,” inICML, 2018

    2018

  13. [21]

    Tent: Fully test-time adaptation by entropy minimiza- tion,

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell, “Tent: Fully test-time adaptation by entropy minimiza- tion,” inICLR, 2021

    2021

  14. [22]

    Casting a bait for offline and online source-free do- main adaptation,

    Shiqi Yang, Yaxing Wang, Luis Herranz, Shangling Jui, and Joost van de Weijer, “Casting a bait for offline and online source-free do- main adaptation,”Comput. Vis. Image Underst., vol. 234, no. C, 2023

    2023

  15. [23]

    Universal source-free domain adaptation,

    Jogendra Nath Kundu, Naveen Venkat, M. V . Rahul, and R. Venkatesh Babu, “Universal source-free domain adaptation,” in CVPR, 2020, pp. 4543–4552

    2020

  16. [24]

    Fixmatch: Simplifying semi-supervised learning with con- sistency and confidence,

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun- Liang Li, “Fixmatch: Simplifying semi-supervised learning with con- sistency and confidence,”NeurIPS, vol. 33, pp. 596–608, 2020

    2020

  17. [25]

    Freematch: Self-adaptive thresh- olding for semi-supervised learning,

    Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, Bernt Schiele, and Xing Xie, “Freematch: Self-adaptive thresh- olding for semi-supervised learning,” inICLR, 2023

    2023

  18. [26]

    Learn- ing to reweight examples for robust deep learning,

    Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun, “Learn- ing to reweight examples for robust deep learning,” inICML. PMLR, 2018, pp. 4334–4343

    2018

  19. [27]

    Cleaning noisy labels by negative ensemble learning for source-free unsupervised do- main adaptation,

    Waqar Ahmed, Pietro Morerio, and Vittorio Murino, “Cleaning noisy labels by negative ensemble learning for source-free unsupervised do- main adaptation,” inWACV, 2022, pp. 356–365

    2022

  20. [28]

    Uncertainty-aware pseudo-label filtering for source-free unsu- pervised domain adaptation,

    Xi Chen, Haosen Yang, Huicong Zhang, Hongxun Yao, and Xiatian Zhu, “Uncertainty-aware pseudo-label filtering for source-free unsu- pervised domain adaptation,”Neurocomputing, vol. 575, 2024

    2024

  21. [29]

    mixup: Beyond empirical risk minimization,

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez- Paz, “mixup: Beyond empirical risk minimization,”ICLR, 2018

    2018

  22. [30]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Jun- suk Choe, and Youngjoon Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” inICCV, 2019

    2019

  23. [31]

    Manifold mixup: Better representations by interpolating hidden states,

    Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio, “Manifold mixup: Better representations by interpolating hidden states,” inICML. PMLR, 2019, pp. 6438–6447

    2019

  24. [32]

    Token- mix: Rethinking image mixing for data augmentation in vision trans- formers,

    Jihao Liu, Boxiao Liu, Hang Zhou, Hongsheng Li, and Yu Liu, “Token- mix: Rethinking image mixing for data augmentation in vision trans- formers,” inECCV. Springer, 2022, pp. 455–471

    2022

  25. [33]

    Transmix: Attend to mix for vision transformers,

    Jie-Neng Chen, Shuyang Sun, Ju He, Philip HS Torr, Alan Yuille, and Song Bai, “Transmix: Attend to mix for vision transformers,” inCVPR, 2022, pp. 12135–12144

    2022

  26. [34]

    Fixbi: Bridging domain spaces for unsupervised domain adaptation,

    Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang, “Fixbi: Bridging domain spaces for unsupervised domain adaptation,” inCVPR, 2021, pp. 1094–1103

    2021

  27. [35]

    Patch-mix transformer for unsupervised domain adaptation: A game perspective,

    Jinjing Zhu, Haotian Bai, and Lin Wang, “Patch-mix transformer for unsupervised domain adaptation: A game perspective,” inCVPR, 2023

    2023

  28. [36]

    Adversarial domain adaptation with do- main mixup,

    Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang, “Adversarial domain adaptation with do- main mixup,” inAAAI, 2020, vol. 34, pp. 6502–6509

    2020

  29. [37]

    Prox- ymix: Proxy-based mixup training with label refinery for source-free domain adaptation,

    Yuhe Ding, Lijun Sheng, Jian Liang, Aihua Zheng, and Ran He, “Prox- ymix: Proxy-based mixup training with label refinery for source-free domain adaptation,”Neural Networks, vol. 167, pp. 92–103, 2023

    2023

  30. [38]

    Understand- ing and improving source-free domain adaptation from a theoretical perspective,

    Yu Mitsuzumi, Akisato Kimura, and Hisashi Kashima, “Understand- ing and improving source-free domain adaptation from a theoretical perspective,” inCVPR, 2024, pp. 28515–28524

    2024

  31. [39]

    Sf(da)2: Source-free domain adaptation through the lens of data aug- mentation,

    Uiwon Hwang, Jonghyun Lee, Juhyeon Shin, and Sungroh Yoon, “Sf(da)2: Source-free domain adaptation through the lens of data aug- mentation,” inICLR, 2024

    2024

  32. [40]

    Im- proved baselines with momentum contrastive learning,

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He, “Im- proved baselines with momentum contrastive learning,”arXiv preprint arXiv:2003.04297, 2020

    2003 arXiv

  33. [41]

    Deeper, broader and artier domain generalization,

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales, “Deeper, broader and artier domain generalization,” inICCV, 2017

    2017

  34. [42]

    Visda: The visual domain adaptation chal- lenge,

    Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko, “Visda: The visual domain adaptation chal- lenge,”arXiv preprint arXiv:1710.06924, 2017

    2017 arXiv

  35. [43]

    Moment matching for multi-source domain adaptation,

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang, “Moment matching for multi-source domain adaptation,” in CVPR, 2019, pp. 1406–1415

    2019

  36. [44]

    Domain separation networks,

    Konstantinos Bousmalis, George Trigeorgis, Nathan Silberman, Dilip Krishnan, and Dumitru Erhan, “Domain separation networks,” in NeurIPS, 2016, vol. 29, p. 343–351

    2016

  37. [45]

    Unsupervised multi-target domain adaptation: An information theoretic approach,

    Behnam Gholami, Pritish Sahu, Ognjen Rudovic, Konstantinos Bous- malis, and Vladimir Pavlovic, “Unsupervised multi-target domain adaptation: An information theoretic approach,”IEEE Transactions on Image Processing, vol. 29, pp. 3993–4002, 2020

    2020

  38. [46]

    Knowledge distilla- tion methods for efficient unsupervised adaptation across multiple do- mains,

    Le Thanh Nguyen-Meidine, Atif Belal, Madhu Kiran, Jose Dolz, Louis-Antoine Blais-Morin, and Eric Granger, “Knowledge distilla- tion methods for efficient unsupervised adaptation across multiple do- mains,”Image and Vision Computing, vol. 108, pp. 104096, 2021

    2021

  39. [47]

    Sliced wasserstein discrepancy for unsupervised domain adap- tation,

    Chen-Yu Lee, Tanmay Batra, Mohammad Haris Baig, and Daniel Ul- bricht, “Sliced wasserstein discrepancy for unsupervised domain adap- tation,” inCVPR, 2019, pp. 10285–10295

    2019

  40. [48]

    Contrastive adaptation network for unsupervised domain adaptation,

    Guoliang Kang, Lu Jiang, Yi Yang, and Alexander G Hauptmann, “Contrastive adaptation network for unsupervised domain adaptation,” inCVPR, 2019, pp. 4893–4902

    2019

  41. [49]

    Model adaptation: Unsupervised domain adaptation without source data,

    Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu, “Model adaptation: Unsupervised domain adaptation without source data,” inCVPR, 2020, pp. 9641–9650. 6

    2020

Pith tools

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