Pith. sign in

REVIEW 3 major objections 4 minor 33 references

Effective Dual-Region Augmentation for Reduced Reliance on Large Amounts of Labeled Data

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims that separately perturbing foreground and background regions of training images improves model generalization enough to reduce reliance on large labeled datasets.

desk verdict A moderately useful augmentation trick with decent results, but the headline claim about reducing label reliance is never actually tested. read the letter →

arxiv 2504.13077 v2 pith:E6A3BVHD submitted 2025-04-17 cs.CV

Prasanna Reddy Pulakurthi , Majid Rabbani , Celso M. de Melo , Sohail A. Dianat , Raghuveer M. Rao This is my paper · ORCID
classification cs.CV
keywords DataAugmentationSource-FreeDomainAdaptationPersonRe-IdentificationForegroundPatchNoiseBackgroundShufflingSalientObjectDetectionPACSbenchmarkRobustGeneralization
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 sets out to show that a training-time augmentation can shrink the need for large labeled datasets by treating the foreground and background of an image differently. It applies random Gaussian noise patches to the foreground and randomly shuffles background patches, guided by a saliency mask, and then recombines the two regions. On source-free domain adaptation over the PACS benchmark, the combined augmentation raises average accuracy over the base contrastive method in both single-target and multi-target settings. On two person re-identification benchmarks, it improves Rank-1 and mean average precision over the baseline and over random erasing. The intended upshot is that structured, region-aware augmentation is a scalable alternative to collecting more annotations.

What carries the argument

The load-bearing object is a binary saliency mask produced by a U-shaped salient object detection network (U2-Net). Given that mask, the method (Algorithm 1) selects random patches whose total area is 2–40% of the image, applies Gaussian noise to foreground pixels inside those patches, divides the image into a random 2×2, 4×4, or 8×8 grid, shuffles all patches, and then restores the original foreground pixels using the mask. The recombined image is the augmented training example, and a mixing probability ρ controls how often the augmentation is applied.

What would settle it

Run the single-target PACS experiment again but replace the U2-Net masks with random binary masks matched to the same foreground area; if random masks reproduce the 84.0% average, then the foreground/background distinction is not what carries the gain, and if accuracy falls, mask accuracy is load-bearing.

Watch

Extended reading notes

Core claim

The central claim is that adding Gaussian noise to salient foreground regions and shuffling background patches—while restoring the original foreground via a binary saliency mask—forces models to rely on holistic object features rather than location or background cues, improving cross-domain generalization. In the paper's experiments, integrating this augmentation into a contrastive test-time adaptation method yields 84.0% average accuracy on PACS in the single-target source-free setting and 77.4% in the multi-target setting, compared with 79.4% and 75.4% for the reproduced base method. On Market-1501 and DukeMTMC-reID, the augmentation surpasses both the baseline and random erasing in Rank-1 and mAP across ResNet-18 and EfficientNet-b4 backbones.

Load-bearing premise

The whole pipeline depends on the mask that separates foreground from background being correct; if it labels pixels wrongly, shuffling the background can harm the object or leave background cues intact, and the reported gains could disappear.

Editorial extensions

If this is right

  • If the claim holds, the same augmentation can be dropped into existing source-free adaptation pipelines to improve target accuracy without using any target labels or source data beyond the pretrained model.
  • The largest reported gains on PACS occur on the hardest style shifts (Photo→Sketch and Art-Painting→Sketch), suggesting the method may be especially useful when the target domain is highly stylized.
  • For person re-identification, the method beats random erasing, indicating that structured foreground and background perturbations add robustness against occlusions and background bias beyond generic erasing.
  • Using the augmentation on only a fraction of minibatch samples (ρ=0.8 for single-target, 0.5 for multi-target) means the technique adds no new data collection cost and is cheap to integrate.

Reading between the lines

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

  • An untested corollary is that the gains should scale with mask quality: if U2-Net mislabels foreground pixels, the augmentation could corrupt the object or leave background cues intact, so replacing U2-Net with ground-truth masks should either sharpen or shrink the reported gains.
  • The paper's patch shuffling destroys spatial layout, which may transfer poorly to tasks where background is informative (e.g., scene or place recognition); testing on such tasks would delimit the method's scope.
  • Because the method already outperforms random erasing in ReID, a natural next experiment is combining it with random erasing or with style-transfer augmentations to see whether the benefits are complementary or overlapping.
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 proposes a dual-region data augmentation method for image classification, source-free domain adaptation (SFDA), and person re-identification (ReID). The method uses a U2-Net saliency mask to separate foreground and background, then applies Gaussian noise patches to the foreground and patch shuffling to the background, and recombines the regions. The authors integrate this augmentation into the AdaContrast SFDA framework and evaluate on PACS, reporting 84.0% average accuracy in the single-target setting and 77.4% in the multi-target setting, compared to reproduced AdaContrast baselines of 79.4% and 75.4%, respectively. For ReID, they report improvements over baselines and Random Erasing on Market-1501 and DukeMTMC-reID with ResNet-18 and EfficientNet-b4 backbones. The paper's stated central contribution is 'reduced reliance on large amounts of labeled data,' but all experiments use full labeled training sets, and no experiment varies the amount of labeled data. The results are single-run with no variance estimates or significance tests.

Significance. If the label-efficiency claim were substantiated, the paper would offer a simple, code-available augmentation with practical value for reducing annotation cost. The idea of jointly perturbing foreground and background with an external saliency model is plausible and the ablation (Table 3) suggests both components contribute. However, the current evidence is limited: the headline claim about reduced reliance on labeled data is untested, and the reported gains are not statistically characterized, with some differences being very small or in the wrong direction on individual metrics. The dependence on U2-Net mask accuracy is also unexamined. Reproducibility is aided by the public code release, which is a strength.

major comments (3)
  1. [Abstract, §1, §6] The paper's stated central contribution—'reduced reliance on large amounts of labeled data' (Abstract, also Introduction and Conclusion)—is never tested. All experiments in §4.1 and §4.2 use the full labeled datasets: the complete PACS source domains (1,670–3,929 images per domain) and the full Market-1501 and DukeMTMC-reID training sets (12,936 images/751 identities and 16,522 images/702 identities). No experiment varies the number of labeled source images, labeled identities, or labeled examples. Consequently, the results establish only that the augmentation improves accuracy under a fixed full-size label budget; they do not support the claimed reduction in reliance on labeled data. The authors should either add experiments with subsampled training data (e.g., 10%, 25%, 50%, 100% of source images or ReID identities) or revise the title, abstract, and conclusion to claim only accuracy/robustness improvement at a fixed label budget.
  2. [§5, Tables 1–4] All reported accuracies are from a single training run per configuration; no standard deviations, multiple seeds, or significance tests are provided. Several reported differences are within typical run-to-run noise for deep models. For example, Table 4 shows Ours at 71.38% vs Random Erasing at 71.11% Rank-1 on Market-1501 with ResNet-18, and Ours is worse than Random Erasing on Rank-5 (87.11 vs 87.23) and Rank-10 (91.95 vs 92.01). Table 1 shows Ours at 98.4% vs AdaContrast at 98.7% on A→P. The Abstract's claim of 'significant accuracy improvements' and the caption of Table 4 claiming the method 'consistently outperforms' are therefore not statistically supported. The authors should report mean±std over multiple random seeds and, where feasible, paired significance tests.
  3. [§3, Algorithm 1] The method's core assumption is that the U2-Net binary mask correctly identifies the foreground for every image, since Algorithm 1 (lines 15–21) restores the original foreground pixels using the condition M(i,j)=1. If the mask is inaccurate, foreground pixels may be discarded or background pixels retained, distorting the object or creating spurious cues. The paper provides no evaluation of mask quality on the PACS domains or on pedestrian images from Market-1501/DukeMTMC, and no sensitivity analysis with respect to the threshold used to binarize the U2-Net probability map. The authors should report mask-quality statistics (e.g., IoU against ground-truth saliency or a manual error rate on a sample) and/or an ablation with an imperfect-mask simulation to demonstrate graceful degradation.
minor comments (4)
  1. [Algorithm 1, lines 5–6] The selection of k random patches and the area approximation 'P Area(Pi)≈A' are underspecified: the patch-size set P is defined for the background shuffling step only, and the tolerance for the area match is not given. Clarify the notation, e.g., A ~ Area(I) × U(0.02, 0.40), and specify the patch sizes used for the foreground noise patches.
  2. [Table 4] For Market-1501 with ResNet-18, the proposed method is not better than Random Erasing on Rank-5 (87.11 vs 87.23) and Rank-10 (91.95 vs 92.01); the text states the method 'surpasses' Random Erasing, which is not accurate for those metrics. Please report these cases accurately.
  3. [Throughout] There are several notation and typographical issues, e.g., 'U2-Net23' should be 'U2-Net [23]', 'A←U(a_l,a_h) Area(I)' is unclear, and 'This augmentation-' contains a broken hyphen. A careful copyedit is recommended.
  4. [§4.1] The paper says it follows the evaluation protocols of NEL [25], but NEL targets noisy-label SFDA; clarify how the single-target and multi-target protocols are defined (e.g., which source-target combinations, whether the same backbone and training schedule are used for all baselines).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the proposed augmentation is an empirical method evaluated against external benchmarks, with no derivation that reduces to its own inputs.

full rationale

The paper makes no formal derivation or prediction that could reduce to fitted inputs. The method is a data augmentation pipeline (Algorithm 1) that combines U2-Net-based foreground masks with Gaussian noise patches and background patch shuffling, and the claimed contribution is empirical: higher accuracy on PACS SFDA and on Market-1501/DukeMTMC-reID compared with baselines such as AdaContrast and Random Erasing. These comparisons are against external benchmarks and published or publicly available baselines, and the reproduced AdaContrast results are explicitly checked against published benchmarks. The only self-citations are to prior GAN and domain-adaptation works by the same authors, and none of those citations is load-bearing for the central claim; the paper does not invoke a uniqueness theorem or a self-cited method as the justification for its design. The stated goal of 'reduced reliance on large amounts of labeled data' is not directly tested by experiments that vary the amount of labeled data, but that is a missing-experiment or scope issue, not circularity. No equation or construction in the paper equates the output with an input, and no fitted parameter is renamed as a prediction. Therefore the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's results rest on two external tools (U2-Net masks and AdaContrast) and several hand-chosen hyperparameters; no new physical or architectural entities are introduced.

free parameters (3)
  • mixing probability rho = 0.8 for single-target SFDA; 0.5 for multi-target SFDA and ReID
    Probability of applying the augmentation; tuned per task with no sensitivity analysis reported (Sections 4.1 and 4.2).
  • noise area ratio range [al, ah] = [0.02, 0.40] of image area
    Uniformly sampled area covered by foreground noise patches; chosen by authors, no ablation across values.
  • patch size set P = {14, 28, 56, 112} for 224x224 images
    Discrete patch sizes for background shuffling; chosen by authors without sensitivity analysis.
assumptions (4)
  • domain assumption U2-Net provides accurate binary saliency masks that separate foreground from background in PACS and ReID images.
    The entire dual-region pipeline depends on mask quality; mask errors would corrupt the augmented image (Algorithm 1, lines 15 to 21).
  • domain assumption Gaussian noise patches and background patch shuffling preserve the class label and identity of the image.
    If these transformations alter semantic content, the augmented images provide incorrect supervision; no label verification is performed.
  • domain assumption The reproduced AdaContrast baseline faithfully represents the original method's performance on PACS.
    The authors generated this baseline themselves using the public codebase and state that it surpasses previously reported best results in NEL (Section 4.1).
  • domain assumption The chosen training hyperparameters (SGD, learning rates, epochs) are sufficient for fair comparison across methods.
    No hyperparameter search is reported for the baselines, and the same settings are used across methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Effective Dual-Region Augmentation for Reduced Reliance on Large Amounts of Labeled Data." pith.science (2026). https://pith.science/paper/E6A3BVHD

@misc{pith2026250413077,
  author       = {Pith},
  title        = {Pith review of: Effective Dual-Region Augmentation for Reduced Reliance on Large Amounts of Labeled Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E6A3BVHD}},
  note         = {Machine review of arXiv:2504.13077}
}
read the original abstract

This paper introduces a novel dual-region augmentation approach designed to reduce reliance on large-scale labeled datasets while improving model robustness and adaptability across diverse computer vision tasks, including source-free domain adaptation (SFDA) and person re-identification (ReID). Our method performs targeted data transformations by applying random noise perturbations to foreground objects and spatially shuffling background patches. This effectively increases the diversity of the training data, improving model robustness and generalization. Evaluations on the PACS dataset for SFDA demonstrate that our augmentation strategy consistently outperforms existing methods, achieving significant accuracy improvements in both single-target and multi-target adaptation settings. By augmenting training data through structured transformations, our method enables model generalization across domains, providing a scalable solution for reducing reliance on manually annotated datasets. Furthermore, experiments on Market-1501 and DukeMTMC-reID datasets validate the effectiveness of our approach for person ReID, surpassing traditional augmentation techniques. The code is available at https://github.com/PrasannaPulakurthi/Foreground-Background-Augmentation

Figures

Figures reproduced from arXiv: 2504.13077 by the authors.

Figure 1
Figure 1. Examples of the proposed augmentation method applied to (a) image classification, (b) domain adaptation, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed data augmentation pipeline. Foreground and background regions are separated using [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 29 canonical work pages

  1. [1]

    D., [Dataset Shift in Machine Learning], The MIT Press (2009)

    Quio˜ nero-Candela, J., Sugiyama, M., Schwaighofer, A., and Lawrence, N. D., [Dataset Shift in Machine Learning], The MIT Press (2009)

    work page 2009

  2. [2]

    Unsupervised domain adaptation by backpropagation,

    Ganin, Y. and Lempitsky, V., “Unsupervised domain adaptation by backpropagation,” in [International conference on machine learning], 1180–1189, PMLR (2015)

    work page 2015

  3. [3]

    Adversarial discriminative domain adaptation,

    Tzeng, E., Hoffman, J., Saenko, K., and Darrell, T., “Adversarial discriminative domain adaptation,” in [Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)], (July 2017)

    work page 2017

  4. [4]

    Conditional adversarial domain adaptation,

    Long, M., Cao, Z., Wang, J., and Jordan, M. I., “Conditional adversarial domain adaptation,”Advances in neural information processing systems31(2018)

    work page 2018

  5. [5]

    Unsupervised domain adaptation using feature aligned maximum classifier discrepancy,

    Pulakurthi, P. R., Dianat, S. A., Rabbani, M., You, S., and Rao, R. M., “Unsupervised domain adaptation using feature aligned maximum classifier discrepancy,” in [Applications of Machine Learning 2022], Zelinski, M. E., Taha, T. M., and Howe, J., eds.,12227, 1222707, International Society for Optics and Photonics, SPIE (2022)

    work page 2022

  6. [6]

    Contrastive test-time adaptation,

    Chen, D., Wang, D., Darrell, T., and Ebrahimi, S., “Contrastive test-time adaptation,” in [Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition], 295–305 (2022)

    work page 2022

  7. [7]

    Generative adversarial nets,

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y., “Generative adversarial nets,”Advances in neural information processing systems27(2014)

    work page 2014

  8. [8]

    Training generative adversarial networks with limited data,

    Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., and Aila, T., “Training generative adversarial networks with limited data,” in [Proc. NeurIPS], (2020)

    work page 2020

Show all 33 references
  1. [9]

    Enhancing gans with mmd neural architecture search, pmish activation function, and adaptive rank decomposition,

    Reddy Pulakurthi, P., Mozaffari, M., Dianat, S. A., Heard, J., Rao, R. M., and Rabbani, M., “Enhancing gans with mmd neural architecture search, pmish activation function, and adaptive rank decomposition,” IEEE Access12, 174222–174244 (2024)

    2024

  2. [10]

    CyCADA: Cycle-consistent adversarial domain adaptation,

    Hoffman, J., Tzeng, E., Park, T., Zhu, J.-Y., Isola, P., Saenko, K., Efros, A., and Darrell, T., “CyCADA: Cycle-consistent adversarial domain adaptation,” in [Proceedings of the 35th International Conference on Machine Learning], Dy, J. and Krause, A., eds.,Proceedings of Mach...

    1989

  3. [11]

    Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis,

    Liu, W., Piao, Z., Min, J., Luo, W., Ma, L., and Gao, S., “Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view synthesis,” in [Proceedings of the IEEE/CVF international conference on computer vision], 5904–5913 (2019)

    2019

  4. [12]

    Enhancing human action recognition with gan-based data augmentation,

    Pulakurthi, P. R., De Melo, C. M., Rao, R., and Rabbani, M., “Enhancing human action recognition with gan-based data augmentation,” in [Synthetic Data for Artificial Intelligence and Machine Learning: Tools, Techniques, and Applications II],13035, 194–204, SPIE (2024)

    2024

  5. [13]

    Auggan: Cross domain adaptation with gan-based data augmentation,

    Huang, S.-W., Lin, C.-T., Chen, S.-P., Wu, Y.-Y., Hsu, P.-H., and Lai, S.-H., “Auggan: Cross domain adaptation with gan-based data augmentation,” in [Proceedings of the European conference on computer vision (ECCV)], 718–731 (2018)

    2018

  6. [14]

    Learnable data augmentation for one-shot unsupervised domain adaptation,

    Carrazco, J. I. D., Morerio, P., Bue, A. D., and Murino, V., “Learnable data augmentation for one-shot unsupervised domain adaptation,” in [34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023], BMVA (2023)

    2023

  7. [15]

    Adversarial and random transformations for robust domain adaptation and generalization,

    Xiao, L., Xu, J., Zhao, D., Shang, E., Zhu, Q., and Dai, B., “Adversarial and random transformations for robust domain adaptation and generalization,”Sensors23(11), 5273 (2023)

    2023

  8. [16]

    Parameter-efficient person re-identification in the 3d space,

    Zheng, Z., Wang, X., Zheng, N., and Yang, Y., “Parameter-efficient person re-identification in the 3d space,” IEEE Transactions on Neural Networks and Learning Systems (TNNLS)(2022). doi:10.1109/TNNLS.2022. 3214834

    2022 doi

  9. [17]

    Joint discriminative and generative learning for person re-identification,

    Zheng, Z., Yang, X., Yu, Z., Zheng, L., Yang, Y., and Kautz, J., “Joint discriminative and generative learning for person re-identification,” in [proceedings of the IEEE/CVF conference on computer vision and pattern recognition], 2138–2147 (2019)

    2019

  10. [18]

    Unlabeled samples generated by gan improve the person re- identification baseline in vitro,

    Zheng, Z., Zheng, L., and Yang, Y., “Unlabeled samples generated by gan improve the person re- identification baseline in vitro,” in [Proceedings of the IEEE international conference on computer vision], 3754–3762 (2017)

    2017

  11. [19]

    An improved cyclegan for data augmentation in person re-identification,

    Yang, Z., Shao, J., and Yang, Y., “An improved cyclegan for data augmentation in person re-identification,” Big Data Research34, 100409 (2023)

    2023

  12. [20]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A., “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in [Proceedings of the IEEE international conference on computer vision], 2223–2232 (2017)

    2017

  13. [21]

    Random erasing data augmentation,

    Zhong, Z., Zheng, L., Kang, G., Li, S., and Yang, Y., “Random erasing data augmentation,” in [Proceedings of the AAAI conference on artificial intelligence],34(07), 13001–13008 (2020)

    2020

  14. [22]

    A person re-identification data augmentation method with adversarial defense effect,

    Gong, Y., Zeng, Z., Chen, L., Luo, Y., Weng, B., and Ye, F., “A person re-identification data augmentation method with adversarial defense effect,”arXiv preprint arXiv:2101.08783(2021)

    2021 arXiv

  15. [23]

    U2-net: Going deeper with nested u-structure for salient object detection,

    Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O., and Jagersand, M., “U2-net: Going deeper with nested u-structure for salient object detection,”Pattern Recognition106, 107404 (2020)

    2020

  16. [24]

    Deeper, broader and artier domain generalization,

    Li, D., Yang, Y., Song, Y.-Z., and Hospedales, T. M., “Deeper, broader and artier domain generalization,” in [Proceedings of the IEEE international conference on computer vision], 5542–5550 (2017)

    2017

  17. [25]

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

    Ahmed, W., Morerio, P., and Murino, V., “Cleaning noisy labels by negative ensemble learning for source- free unsupervised domain adaptation,” in [2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)], 356–365 (2022)

    2022

  18. [26]

    Deep residual learning for image recognition,

    He, K., Zhang, X., Ren, S., and Sun, J., “Deep residual learning for image recognition,” in [Proceedings of the IEEE conference on computer vision and pattern recognition], 770–778 (2016)

    2016

  19. [27]

    Scalable person re-identification: A benchmark,

    Zheng, L., Shen, L., Tian, L., Wang, S., Wang, J., and Tian, Q., “Scalable person re-identification: A benchmark,” in [Proceedings of the IEEE International Conference on Computer Vision (ICCV)], (December 2015)

    2015

  20. [28]

    Performance measures and a data set for multi-target, multi-camera tracking,

    Ristani, E., Solera, F., Zou, R., Cucchiara, R., and Tomasi, C., “Performance measures and a data set for multi-target, multi-camera tracking,” in [European conference on computer vision], 17–35, Springer (2016)

    2016

  21. [29]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    Tan, M. and Le, Q., “Efficientnet: Rethinking model scaling for convolutional neural networks,” in [Inter- national conference on machine learning], 6105–6114, PMLR (2019)

    2019

  22. [30]

    Domain separation networks,

    Bousmalis, K., Trigeorgis, G., Silberman, N., Krishnan, D., and Erhan, D., “Domain separation networks,” in [Proceedings of the 30th International Conference on Neural Information Processing Systems],NIPS’16, 343–351, Curran Associates Inc., Red Hook, NY, USA (2016)

    2016

  23. [31]

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

    Gholami, B., Sahu, P., Rudovic, O., Bousmalis, K., and Pavlovic, V., “Unsupervised multi-target domain adaptation: An information theoretic approach,”IEEE Transactions on Image Processing29, 3993–4002 (2020)

    2020

  24. [32]

    Knowledge distillation methods for efficient unsupervised adaptation across multiple domains,

    Nguyen-Meidine, L. T., Belal, A., Kiran, M., Dolz, J., Blais-Morin, L.-A., and Granger, E., “Knowledge distillation methods for efficient unsupervised adaptation across multiple domains,”Image and Vision Com- puting108, 104096 (2021)

    2021

  25. [33]

    Improved baselines with momentum contrastive learning,

    Chen, X., Fan, H., Girshick, R., and He, K., “Improved baselines with momentum contrastive learning,” arXiv preprint arXiv:2003.04297(2020)

    2020 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.