Pith. sign in

REVIEW 5 major objections 6 minor 71 references

Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization

T0 review · 5 major / 6 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 treating augmented copies of a single dataset as separate pseudo-domains lets multi-source domain generalization (MDG) algorithms outperform single-source baselines, and that these pseudo-domains can rival real…

desk verdict Useful benchmark and empirical map, but the headline outperformance claim is unsupported because the PACS gains come from test-domain-conditioned training. read the letter →

arxiv 2505.23173 v1 pith:SPS3BYR5 submitted 2025-05-29 cs.LG cs.CV

Shohei Enomoto This is my paper
classification cs.LGcs.CV
keywords domaingeneralizationsingle-sourcemulti-sourcepseudo-domaindataaugmentationstyletransferdomain-adversarialtrainingcorrelationalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that the powerful algorithms developed for multi-source domain generalization (MDG) can be transplanted into the much cheaper single-source setting, where only one training distribution exists. Its recipe, PMDG, turns each mini-batch into several 'pseudo-domains' by applying distinct transformations (style transfer, cartoonization, edge detection, or strong augmentations) and then trains with a standard MDG loss over these synthetic domains. The reported experiments across four domain-shift datasets and two backbones indicate that this recipe beats the best single-source baselines, and that with enough data pseudo-domains can match or surpass training on genuinely multi-domain datasets. If correct, this would let practitioners skip costly multi-domain data collection and still use the sophisticated tools of MDG research.

What carries the argument

The engine of the method is the pseudo-domain, defined by the transformation rule B_k = O_k(B) that converts one training mini-batch into K labeled domain copies. The paper evaluates two transformation families as pseudo-domain generators: style transforms (AdaIN style transfer, CartoonGAN, edge detection) that aim to mimic the Photo/Art/Cartoon/Sketch split of PACS, and data augmentations (IPMix, RandConv, TrivialAugment, AugMix, MixUp, CutMix, RandAugment) that alter low-level appearance. These copies are fed to any MDG algorithm's loss, such as domain-adversarial training or correlation alignment, which is what makes the bridge between the single-source and multi-source paradigms.

What would settle it

A direct test would compare PMDG against the same training pipeline in which all pseudo-domain copies are pooled into one domain, so the MDG loss sees only a single distribution. If the pooled version matches PMDG's accuracy, then the domain labels provide no benefit and the reported gains are just augmentation. A complementary check is to measure the feature-space distance between the source and each transformed copy across batches; gains should vanish when these distances are comparable to the noise between batches of the source itself.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the 'domain' variable in multi-source domain generalization does not have to come from natural data collection: it can be manufactured by transforming a single source. PMDG defines K pseudo-domains by applying transformations O_1,...,O_K to the same mini-batch, giving B_k = O_k(B), and then minimizes an MDG objective over these K copies. The paper's experiments show that this configuration outperforms existing single-source DG methods (the best average result on its four datasets uses the SD algorithm with two IPMix pseudo-domains), that MDG algorithm rankings transfer to the pseudo-domain setting via a positive correlation, and that under equal sample budgets PMDG can exceed the accuracy of true multi-domain training in several source-domain configurations. A further claim is that future single-source DG progress should come from designing better pseudo-domain transformations rather than from inventing new learning algorithms.

Load-bearing premise

The entire approach depends on the assumption that the transformed copy of a mini-batch is a stable, distinct domain rather than another random view of the same domain; if transformations are weak or stochastic, the pseudo-domain structure is an illusion and MDG losses have nothing meaningful to align.

Editorial extensions

If this is right

  • MDG algorithms become plug-in single-source trainers, so a large body of multi-source research transfers directly to single-source problems.
  • MDG algorithm rankings in the pseudo-domain setting correlate with their rankings on real multi-domain data, so pseudo-domains could serve as a cheap proxy for algorithm selection.
  • Style-based pseudo-domains give large gains on PACS but not on other datasets, so benchmark diversity is necessary before declaring any transformation strategy a winner.
  • Repeated use of one strong transformation (IPMix) beats mixing many different transformations, implying that the consistency of the pseudo-domain identity matters more than transformation diversity.

Reading between the lines

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

  • A testable extension is that the same pseudo-domain recipe should carry over to non-image modalities, such as audio or text, where style-like transforms (voice conversion, paraphrase, synthetic noise) could generate pseudo-domains; the paper does not address this.
  • The positive MDG-PMDG correlation suggests that pseudo-domain evaluation could reduce the cost of MDG algorithm development by replacing real multi-domain benchmarking with synthetic-domain runs.
  • If pseudo-domains genuinely substitute for natural domains, the practical goal of domain generalization shifts from collecting many domains to collecting one rich source plus a good transformation set; the paper hints at this but does not develop it.
  • The uniform failure of MLDG across all transformations is a useful anomaly: understanding why a meta-learning objective breaks on pseudo-domains could sharpen the boundary between real and synthetic domain structure.
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

5 major / 6 minor

Summary. The paper proposes Pseudo Multi-source Domain Generalization (PMDG), a framework that generates multiple pseudo-domains from a single source domain via style transfer and data augmentation, then applies existing multi-source domain generalization (MDG) algorithms to these pseudo-domains. The authors introduce PseudoDomainBed, a modified DomainBed benchmark, and report experiments on PACS, VLCS, OfficeHome, TerraIncognita, and ImageNet with ResNet50 and ViT backbones. The central claims are that PMDG outperforms existing single-source DG (SDG) baselines, that MDG algorithm rankings correlate between MDG and PMDG settings, and that pseudo-domains can match or exceed real multi-domain performance with sufficient data.

Significance. If established, the framework would be a practical bridge between SDG and MDG, allowing sophisticated MDG algorithms to be used when only one domain is available. The release of PseudoDomainBed and the systematic evaluation across many transformation/algorithm combinations are useful contributions to empirical DG research. However, the headline performance claims are currently undermined by a test-domain-dependent training protocol and by post hoc selection of pseudo-domain combinations; the evidence for the 'sufficient data' claim is also weak. The paper's value would increase substantially if the experiments were rerun with fixed models and pre-specified transformation selection.

major comments (5)
  1. [Section 6.2, Table 1 (dagger rows)] The dagger protocol in Table 1 conditions the training configuration on the test domain: ST is excluded when testing on Art, CG when testing on Cartoon, and ED when testing on Sketch. Consequently, the reported PACS averages for rows such as ERM Org+ST+ED+CG+IM+IM† (69.9) and RIDG Org+ST+ED+CG+IM+IM† (71.8) do not correspond to any single fixed model or method; they aggregate four different training configurations selected using knowledge of the target domain. This leaks target-domain identity into the training protocol and makes the comparison with fixed SDG baselines (IPMix, TrivialAugment, etc.) invalid. Without the dagger exclusion, the best fixed-configuration PMDG result on PACS is SD Org+IM+IM at 64.1, which is below IPMix (65.9±0.3), and the average gain over IPMix (55.9 vs 55.2) is within one standard error. The central claim that 'PMDG outperforms existing SDG methods' is therefore not supported by the results as presented.
  2. [Section 4.1.3 and Section 6.2] The paper does not specify how the pseudo-domain combinations reported in Table 1 (Org+IM+IM and Org+ST+ED+CG+IM+IM) were selected. Section 4.1.3 states only that 'we take an empirical approach' with 'limited understanding of optimal transformation count and inter-transformation interactions.' If these combinations were chosen after examining results across the many configurations in Table 3 or Figure 3, then the reported 'superior performance' is partly a selection artifact rather than an independently predicted outcome. A pre-registered selection rule, ideally using only training-domain validation, is necessary to support the claim that PMDG outperforms SDG baselines.
  3. [Section 6.4, Figure 5 vs. Abstract] The abstract and Section 9 claim that 'with sufficient training data' pseudo-domains can match or exceed actual multi-domain performance, but Figure 5 does not support this claim. The controlled comparison shows that PMDG's performance does not consistently improve with dataset size and varies strongly with the choice of source domain; in several panels (e.g., VLCS test domain L, PACS test domain C, OfficeHome test domain A) MDG outperforms PMDG at all sample sizes shown. The claim of a 'sufficient data' regime is not established, and the sentence in Section 6.4 that PMDG 'can achieve superior performance with specific source domains' indicates that the effect is source-domain-dependent rather than driven primarily by data quantity.
  4. [Section 4.1.4 and Section 8] The framework's premise, stated in Eq. (3), is that each transformed mini-batch B_k = O_k(B) constitutes a distinct pseudo-domain. However, for stochastic augmentations such as IPMix or AugMix, the transformation changes on every mini-batch, so the 'domain' seen by the MDG algorithm (e.g., DANN, CORAL, MMD) is resampled each step rather than being a fixed distribution. MDG losses that align domain distributions are designed for fixed domains; applying them to ever-changing random augmentations is not obviously justified. Section 8 itself concedes that weakly transformed data may remain substantially similar to the source distribution. The paper should either provide a formal or empirical justification that the MDG objective remains meaningful under this construction, or temper the claim of 'applying MDG algorithms' and present PMDG primarily as an augmentation framework.
  5. [Section 6, Table 1] The headline comparisons are reported without confidence intervals or significance tests on the aggregated 'Avg' column. For example, the best PMDG average (SD Org+IM+IM, 55.9) differs from the best baseline (IPMix, 55.2) by 0.7 percentage points, which is within the typical per-dataset standard errors reported elsewhere in the table. The paper should report the standard error for the averages, or a paired significance test across datasets, before claiming that PMDG outperforms SDG baselines.
minor comments (6)
  1. [Table 2] The header 'IN- C' appears to be a typo for 'IN-C', and 'Stylied-IN' should be 'Stylized-IN'; also, no standard errors are reported for the ImageNet results, making it impossible to assess the reliability of the 0.30-point OOD average improvement.
  2. [Table 3] Several entries in the supplementary table lack standard errors (e.g., ERM Org+ST+ED+CT † and SD Org+ST+ED+CT †), so those results cannot be compared meaningfully with the rest of the table.
  3. [Section 5.1.2] The description of dataset-level versus mini-batch-level transformations is useful, but the paper never gives a complete mapping of which transformations fall into each category; a small table would improve reproducibility.
  4. [Section 9] The conclusion uses 'Single-Domain Generalization' while the rest of the paper uses 'Single-source Domain Generalization'; the terminology should be consistent.
  5. [Figure 3] The heatmap values in Figure 3 are very difficult to read at the printed size; enlarging the figure or providing the numerical values in a table would help readers evaluate the per-algorithm gains.
  6. [Section 8] The limitation section is candid, but it directly contradicts the premise in Section 4.1.4 that all transformations create distinct domains; the discussion in Section 7 should acknowledge this tension explicitly rather than presenting it only as a future-work item.

Circularity Check

1 steps flagged · score 6.0 of 10

PACS outperformance claim is target-conditioned by the dagger protocol, not a fixed-method prediction.

  1. fitted input called prediction [Table 1 footnote (Section 6.2); Section 4.1.1; Section 6.2]
    "† indicates exclusion of domain-specific transformations during training: ST is excluded when testing on Art domain, CG for Cartoon domain, and ED for Sketch domain. ... Inspired by the PACS dataset, we propose three transformations to recreate its constituent domains. The first transformation is AdaIN style transfer ... used for creating art-style images. The second is CartoonGAN ... The third is Edge Detection ... used to generate sketch-style images."

    The dagger protocol defines, for each PACS target, a training configuration that removes exactly the transformation designed to mimic that target's style (ST for Art, CG for Cartoon, ED for Sketch). Consequently, each reported PACS per-domain accuracy is produced by a different model whose pseudo-domain set is selected using knowledge of the test domain. The reported PACS average is thus the mean of four target-conditioned training configurations, not the accuracy of any single fixed PMDG method. Comparing this constructed average to fixed SDG baselines (IPMix, TrivialAugment, etc.) makes the headline claim 'PMDG outperforms existing SDG methods' an artifact of test-domain information rather than a prediction.

full rationale

The paper is primarily an empirical study with no formal derivation, so circularity must be assessed through the evaluation protocol. The one concrete reduction I can exhibit is the Table 1 dagger protocol for PACS, which is load-bearing for the central 'PMDG outperforms existing SDG methods' claim. For each PACS target, the training set excludes the pseudo-domain transformation that was explicitly designed to recreate that target's style; hence the reported PACS numbers are not generated by any single PMDG method but by four different methods selected using the target identity. This is a fitted-input-called-prediction pattern: the test-domain identity is used to configure training, and the resulting accuracy is then reported as method performance. The same dagger appears in the ViT table and Supplemental Table 3, so the most striking PACS gains all depend on this target-conditioned protocol. By contrast, the VLCS, OfficeHome, TerraIncognita, and ImageNet results use fixed pseudo-domain configurations and are not circular in the same way. There are no load-bearing self-citations, and the Section 8 limitation about weakly transformed data being similar to the source is an acknowledged assumption rather than a circular step. Overall, the circularity is partial but real: the PACS-based outperformance prediction is constructed, not independently predicted.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on treating synthetic transformations as real domains (a domain assumption), on per-batch stochastic augmentations behaving as fixed domains (another domain assumption), and on a validation rule borrowed from DomainBed. The transformation set and the dagger exclusions are post-hoc choices that function as fitted parameters.

free parameters (3)
  • Number of pseudo-domains K = 2 or 6
    The paper tests combinations with different numbers of transformations (e.g., Org+IM+IM has K=2 fake domains, Org+ST+ED+CG+IM+IM has K=5 fake domains plus original). Results vary with K, and K is chosen by hand after observing test performance.
  • Pseudo-domain transformation set = e.g., Org+IM+IM, Org+ST+ED+CG+IM+IM
    The specific set of transformations is selected after evaluating many combinations on the same test datasets (Tables 1 and 3), making it a fitted choice rather than a priori.
  • Test-domain-dependent dagger exclusion = applied only when testing on Art/Cartoon/Sketch
    The dagger rule removes StyleTransfer when the test domain is Art, CartoonGAN for Cartoon, and EdgeDetection for Sketch. This is a post-hoc, target-dependent protocol choice that changes the training data per test domain.
assumptions (4)
  • domain assumption Pseudo-domains are stable, distinct domain distributions suitable for MDG algorithms
    Introduced in Section 4.1 and used throughout; the Limitation section admits this assumption is questionable for weak transformations.
  • domain assumption Stochastic per-batch transformations can be treated as fixed domains for MDG losses
    Section 4.2 applies MDG losses to per-batch pseudo-domains without justification; MDG algorithms assume fixed domain distributions.
  • domain assumption Training-domain validation is a reliable model selection criterion in the single-source pseudo-domain setting
    Section 5.1.1 follows DomainBed's training-domain validation, but with one real domain the validation set is drawn from the same distribution as training and may not predict OOD performance.
  • ad hoc to paper StyleTransfer, CartoonGAN, and EdgeDetection create domains analogous to PACS Art, Cartoon, and Sketch
    Section 4.1.1 says the choices are 'Inspired by the PACS dataset'; this analogy justifies the pseudo-domain set but is not independently validated.
invented entities (1)
  • pseudo-domain
    purpose: A synthetic training domain created by applying a transformation to source images, used to run MDG algorithms in a single-source setting.
    The paper introduces this construct as the core of PMDG, but provides no falsifiable handle outside its own benchmark; its validity is the central assumption to be tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization." pith.science (2026). https://pith.science/paper/SPS3BYR5

@misc{pith2026250523173,
  author       = {Pith},
  title        = {Pith review of: Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPS3BYR5}},
  note         = {Machine review of arXiv:2505.23173}
}
read the original abstract

Deep learning models often struggle to maintain performance when deployed on data distributions different from their training data, particularly in real-world applications where environmental conditions frequently change. While Multi-source Domain Generalization (MDG) has shown promise in addressing this challenge by leveraging multiple source domains during training, its practical application is limited by the significant costs and difficulties associated with creating multi-domain datasets. To address this limitation, we propose Pseudo Multi-source Domain Generalization (PMDG), a novel framework that enables the application of sophisticated MDG algorithms in more practical Single-source Domain Generalization (SDG) settings. PMDG generates multiple pseudo-domains from a single source domain through style transfer and data augmentation techniques, creating a synthetic multi-domain dataset that can be used with existing MDG algorithms. Through extensive experiments with PseudoDomainBed, our modified version of the DomainBed benchmark, we analyze the effectiveness of PMDG across multiple datasets and architectures. Our analysis reveals several key findings, including a positive correlation between MDG and PMDG performance and the potential of pseudo-domains to match or exceed actual multi-domain performance with sufficient data. These comprehensive empirical results provide valuable insights for future research in domain generalization. Our code is available at https://github.com/s-enmt/PseudoDomainBed.

Figures

Figures reproduced from arXiv: 2505.23173 by the authors.

Figure 1
Figure 1. Overview of PMDG framework. PMDG applies mul [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of the transformed sample. We performed different types of transformations on the dog images of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Accuracy gains over the ERM baseline without pseudo-domain across different transformation techniques (y-axis) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Accuracy comparison of MDG algorithms across [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Comparison of accuracy between MDG and PMDG settings under equal training data conditions. The x-axis shows [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of Various Image Transformation Techniques. Default Data Augmentation refers to the standard data [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 64 canonical work pages

  1. [1]

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. 2019. Invariant risk minimization. arXiv preprint arXiv:1907.02893 (2019)

    arXiv 2019

  2. [2]

    Sara Beery, Grant Van Horn, and Pietro Perona. 2018. Recognition in terra incognita. In European Conference on Computer Vision (ECCV)

    work page 2018

  3. [3]

    Gilles Blanchard, Aniket Anand Deshmukh, Urun Dogan, Gyemin Lee, and Clayton Scott. 2021. Domain generalization by marginal transfer learning.Journal of machine learning research (JMLR) (2021)

    work page 2021

  4. [4]

    Fabio M Carlucci, Antonio D’Innocente, Silvia Bucci, Barbara Caputo, and Tatiana Tommasi. 2019. Domain generalization by solving jigsaw puzzles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2019

  5. [5]

    Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. 2021. SWAD: Domain Generalization by Seeking Flat Minima. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2021

  6. [6]

    Liang Chen, Yong Zhang, Yibing Song, Anton van den Hengel, and Lingqiao Liu

    work page

  7. [7]

    Yang Chen, Yu-Kun Lai, and Yong-Jin Liu. 2018. Cartoongan: Generative adver- sarial networks for photo cartoonization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2018

  8. [8]

    Seokeon Choi, Debasmit Das, Sungha Choi, Seunghan Yang, Hyunsin Park, and Sungrack Yun. 2023. Progressive random convolutions for single domain gen- eralization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2023

Show all 71 references
  1. [9]

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. 2020. Randaug- ment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    2020

  2. [10]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recogn...

    2021

  3. [11]

    Cian Eastwood, Alexander Robey, Shashank Singh, Julius Von Kügelgen, Hamed Hassani, George J Pappas, and Bernhard Schölkopf. 2022. Probable domain generalization via quantile risk minimization. Advances in Neural Information Processing Systems (NeurIPS) (2022)

    2022

  4. [12]

    Rockmore

    Chen Fang, Ye Xu, and Daniel N. Rockmore. 2013. Unbiased Metric Learning: On the Utilization of Multiple Datasets and Web Images for Softening Bias. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2013

  5. [13]

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. 2016. Domain-adversarial training of neural networks. Journal of machine learning research (JMLR) (2016)

    2016

  6. [14]

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. 2018. ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations (ICLR)

    2018

  7. [15]

    Ishaan Gulrajani and David Lopez-Paz. 2020. In Search of Lost Domain General- ization. In International Conference on Learning Representations (ICLR)

    2020

  8. [16]

    Isabelle Guyon, Vladimir Vapnik, Bernhard Boser, Leon Bottou, and Sara A Solla

  9. [17]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  10. [18]

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. 2021. The many faces of robustness: A critical analysis of out-of-distribution generalization. In IEEE/CVF International Conference on ...

    2021

  11. [19]

    Dan Hendrycks and Thomas Dietterich. 2019. Benchmarking Neural Network Ro- bustness to Common Corruptions and Perturbations. In International Conference on Learning Representations (ICLR)

    2019

  12. [20]

    Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan

    Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. 2020. AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty. In International Conference on Learning Representations (ICLR)

    2020

  13. [21]

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song

  14. [22]

    Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. 2022. Pixmix: Dreamlike pictures comprehensively im- prove safety measures. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  15. [23]

    Xun Huang and Serge Belongie. 2017. Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2017

  16. [24]

    Zhenglin Huang, Xiaoan Bao, Na Zhang, Qingqi Zhang, Xiao Tu, Biao Wu, and Xi Yang. 2023. Ipmix: Label-preserving data augmentation method for training robust classifiers. Advances in Neural Information Processing Systems (NeurIPS) (2023)

    2023

  17. [25]

    Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. 2020. Self-challenging improves cross-domain generalization. In European Conference on Computer Vision (ECCV)

    2020

  18. [26]

    Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. 2022. Style Neophile: Constantly Seeking Novel Styles for Domain Generalization. In IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    2022

  19. [27]

    Daehee Kim, Youngjun Yoo, Seunghyun Park, Jinkyu Kim, and Jaekoo Lee. 2021. Selfreg: Self-supervised contrastive regularization for domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2021

  20. [28]

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. 2021. Out-of- distribution generalization via risk extrapolation (rex). InInternational Conference on Machine Learning (ICML)

    2021

  21. [29]

    Bo Li, Yifei Shen, Jingkang Yang, Yezhen Wang, Jiawei Ren, Tong Che, Jun Zhang, and Ziwei Liu. 2023. Sparse Mixture-of-Experts are Domain Generalizable Learners. In International Conference on Learning Representations (ICLR)

    2023

  22. [30]

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. 2018. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)

    2018

  23. [31]

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. 2017. Deeper, broader and artier domain generalization. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2017

  24. [32]

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. 2018. Domain gener- alization with adversarial feature learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2018

  25. [33]

    Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. 2021. Progressive domain expansion network for single domain generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2021

  26. [34]

    Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. 2018. Deep Domain Generalization via Conditional Invariant Adversarial Networks. In European Conference on Computer Vision (ECCV)

    2018

  27. [35]

    Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, et al. 2020. Shape-Texture Debiased Neural Network Training. In International Conference on Learning Representations (ICLR)

    2020

  28. [36]

    Eric Mintun, Alexander Kirillov, and Saining Xie. 2021. On interaction between augmentations and corruptions in natural corruption robustness. Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  29. [37]

    Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. 2017. Unified deep supervised domain adaptation and generalization. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2017

  30. [38]

    Samuel G Müller and Frank Hutter. 2021. Trivialaugment: Tuning-free yet state- of-the-art data augmentation. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2021

  31. [39]

    Jaemin Na, Heechul Jung, Hyung Jin Chang, and Wonjun Hwang. 2021. Fixbi: Bridging domain spaces for unsupervised domain adaptation. In IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR)

    2021

  32. [40]

    Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. 2021. Reducing domain gap by reducing style bias. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2021

  33. [41]

    Mohammad Pezeshki, Oumar Kaba, Yoshua Bengio, Aaron C Courville, Doina Precup, and Guillaume Lajoie. 2021. Gradient starvation: A learning proclivity in neural networks. Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  34. [42]

    Fengchun Qiao and Xi Peng. 2021. Uncertainty-guided model generalization to unseen domains. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2021

  35. [43]

    Fengchun Qiao, Long Zhao, and Xi Peng. 2020. Learning to learn single do- main generalization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 12556–12565

    2020

  36. [44]

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. 2019. Do imagenet classifiers generalize to imagenet?. In International Conference on Machine Learning (ICML)

    2019

  37. [45]

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al

  38. [46]

    Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. 2020. Distributionally Robust Neural Networks. InInternational Conference on Learning Representations (ICLR)

    2020

  39. [47]

    Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. 2022. Gradient Matching for Domain Generalization. In International Conference on Learning Representations (ICLR) . Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Enomoto

    2022

  40. [48]

    Soria, E

    X. Soria, E. Riba, and A. Sappa. 2020. Dense Extreme Inception Network: Towards a Robust CNN Model for Edge Detection. In IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)

    2020

  41. [49]

    Xavier Soria, Angel Sappa, Patricio Humanante, and Arash Akbarinia. 2023. Dense extreme inception network for edge detection. Pattern Recognition (2023)

    2023

  42. [50]

    Baochen Sun and Kate Saenko. 2016. Deep coral: Correlation alignment for deep domain adaptation. In Computer Vision–ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14

    2016

  43. [51]

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. 2021. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning (ICML)

    2021

  44. [52]

    Shikhar Tuli, Ishita Dasgupta, Erin Grant, and Thomas L Griffiths. 2021. Are Convolutional Neural Networks or Transformers more like human vision?. In 43rd Annual Meeting of the Cognitive Science Society: Comparative Cognition: Animal Minds, CogSci 2021

    2021

  45. [53]

    Puru Vaish, Shunxin Wang, and Nicola Strisciuglio. 2024. Fourier-basis functions to bridge augmentation gap: Rethinking frequency augmentation in image clas- sification. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2024

  46. [54]

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. 2017. Deep hashing network for unsupervised domain adaptation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2017

  47. [55]

    Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. 2018. Generalizing to unseen domains via adversarial data augmentation. Advances in Neural Information Processing Systems (NeurIPS) (2018)

    2018

  48. [56]

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. 2019. Learning Robust Global Representations by Penalizing Local Predictive Power. InAdvances in Neural Information Processing Systems (NeurIPS)

    2019

  49. [57]

    Haotao Wang, Chaowei Xiao, Jean Kossaifi, Zhiding Yu, Anima Anandkumar, and Zhangyang Wang. 2021. AugMax: Adversarial Composition of Random Augmentations for Robust Training. InAdvances in Neural Information Processing Systems (NeurIPS)

    2021

  50. [58]

    Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh

  51. [59]

    Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. 2021. Ro- bust and Generalizable Visual Representation Learning via Random Convolutions. In International Conference on Learning Representations (ICLR)

    2021

  52. [60]

    Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. 2020. Improve unsu- pervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677 (2020)

    2020 arXiv

  53. [61]

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. 2019. Cutmix: Regularization strategy to train strong classifiers with localizable features. InIEEE/CVF International Conference on Computer Vision (ICCV)

    2019

  54. [62]

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2018. mixup: Beyond Empirical Risk Minimization. InInternational Conference on Learn- ing Representations (ICLR)

    2018

  55. [63]

    In IEEE/CVF Inter- national Conference on Computer Vision (ICCV)

    Learning to diversify for single domain generalization. In IEEE/CVF Inter- national Conference on Computer Vision (ICCV)

  56. [64]

    Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. 2020. Maximum-entropy adversarial data augmentation for improved generalization and robustness. Ad- vances in Neural Information Processing Systems (NeurIPS) (2020)

    2020

  57. [65]

    Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. 2020. Learn- ing to generate novel domains for domain generalization. InEuropean Conference on Computer Vision (ECCV)

    2020

  58. [66]

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. 2021. Domain General- ization with MixStyle. In International Conference on Learning Representations (ICLR). Pseudo Multi-Source Domain Generalization: Bridging the Gap Between Single and Multi-Source Domain Generalization Co...

    2021

  59. [68]

    Marvin Zhang, Henrik Marklund, Nikita Dhawan, Abhishek Gupta, Sergey Levine, and Chelsea Finn. 2021. Adaptive risk minimization: Learning to adapt to domain shift. Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  60. [1991]

    Advances in Neural Information Processing Systems (NeurIPS) (1991)

    Structural risk minimization for character recognition. Advances in Neural Information Processing Systems (NeurIPS) (1991)

    1991

  61. [2015]

    International journal of computer vision (IJCV) (2015)

    Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV) (2015)

    2015

  62. [2021]

    In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Natural Adversarial Examples. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

  63. [2023]

    In IEEE/CVF International Conference on Computer Vision (ICCV)

    Domain Generalization via Rationale Invariance. In IEEE/CVF International Conference on Computer Vision (ICCV)

Pith tools

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