Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

General Methods Make Great Domain-specific Foundation Models: A Case-study on Fetal Ultrasound

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Pretraining a standard vision method on 2 million fetal ultrasound images beats much larger natural-image foundation models on fetal tasks.

desk verdict Solid empirical case study, but the prescriptive conclusion about avoiding novel pretraining methods is confounded by the USFM comparison. read the letter →

arxiv 2506.19552 v1 pith:ZJSO4KA2 submitted 2025-06-24 cs.CV cs.AIcs.LG

Jakob Ambsdorf , Asbj{o}rn Munk , Sebastian Llambias , Anders Nymark Christensen , Kamil Mikolaj , Randall Balestriero , Martin Tolsgaard , Aasa Feragen
show 1 more author
Mads Nielsen
This is my paper
classification cs.CVcs.AIcs.LG
keywords self-supervisedlearningfetalultrasoundfoundationmodelsDINOv2domain-specificpretrainingfew-shotsegmentationmedicalimageclassificationvisiontransformers
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 asks whether a medical imaging group should pretrain its own foundation model or reuse a large general-purpose one, and whether new pretraining methods are needed. It answers both by training the well-established DINOv2 self-supervision recipe from scratch on FUS2M, a 2-million-image collection of fetal ultrasound scans from a Danish screening program, yielding the model UltraDINO. On three public fetal ultrasound datasets from Spain, Brazil, and China, UltraDINO outperforms natural-image foundation models trained on up to 142 million images, a general ultrasound foundation model, and supervised baselines on classification, segmentation, and few-shot tasks. The authors conclude that domain-specific pretraining is worth the compute and that mature computer-vision methods can be applied directly, without hyperparameter tuning or methodological novelty, to build strong domain foundation models.

What carries the argument

UltraDINO is built on DINOv2-style self-distillation: a student vision transformer and a teacher vision transformer see different crops and augmentations of the same image; the student must predict the teacher's global [CLS] token and its local masked patch tokens, while the teacher is an exponential moving average of the student, preventing representation collapse. This is a named, well-established mechanism from computer vision. The paper's key move is to apply this exact mechanism, with no hyperparameter tuning, to FUS2M, a 2-million-image fetal ultrasound collection, and to show that the resulting embeddings are more linearly separable for standard fetal planes and that attention heads specialize to anatomical structures, which natural-image baselines do not exhibit. The pretraining data, not a novel loss, is the load-bearing input.

What would settle it

Take the same DINOv2 recipe and pretrain it on a fetal ultrasound corpus of comparable size collected with different scanner hardware outside Denmark, then evaluate on the same three benchmarks; if the performance advantage over natural-image baselines disappears or reverses, the paper's conclusion is specific to the FUS2M composition rather than to fetal ultrasound as a domain.

Watch

Extended reading notes

Core claim

The central claim is that a mature self-supervised method from computer vision, DINOv2, when pretrained from scratch on a sufficiently large, narrow-domain dataset, produces a foundation model superior for that domain to models pretrained on vastly larger natural-image corpora or on broader ultrasound data. UltraDINO, a ViT-B/16 and ViT-S/16 trained for 40 epochs on FUS2M, reaches a Dice score of 86.01 on the four-patient few-shot JNU-IFM segmentation task, versus 83.55 for DINOv2 and 83.05 for USFM, and an F1 of 95.06 in linear-probing classification on the Fetal Planes dataset, versus 76.61 for DINOv2 and 42.78 for USFM. The paper reports that the largest gains appear exactly where labels are scarce, and that larger natural-image pretraining corpora do not keep improving downstream fetal performance. From this the authors argue that domain groups should invest in high-quality domain-specific pretraining data and use established self-supervised learning methods rather than chase novel pretraining methodology.

Load-bearing premise

The central assumption is that the Danish FUS2M corpus, collected with one scanner family in a regional screening program, is representative enough of fetal ultrasound across countries and acquisition protocols that features learned from it transfer to the Spanish, Brazilian, and Chinese evaluation datasets.

Editorial extensions

If this is right

  • Medical imaging groups with access to a large unlabeled domain dataset can expect an off-the-shelf self-supervised recipe to beat transfer from much larger natural-image models on few-shot downstream tasks.
  • For fetal ultrasound, a small ViT-S trained on domain data matches or beats a ViT-B trained on natural images, so scaling up natural-image pretraining is not a substitute for domain-specific pretraining.
  • Existing general ultrasound foundation models trained with masked-image modeling on broad, unbalanced ultrasound data are not the best starting point for a narrow domain like fetal ultrasound.
  • Few-shot and label-efficiency evaluations, not full-dataset finetuning, are where domain-specific pretraining shows its advantage, since full-data performance can saturate at similar levels across pretraining methods.

Reading between the lines

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

  • The same recipe could plausibly transfer to other narrow imaging domains, such as cardiac ultrasound, retinal OCT, or pathology, whenever a comparable corpus of unlabeled clinical images exists, since the mechanism that carries the result is domain data plus a mature self-supervised method, not a fetal-specific design.
  • If scanner- and protocol-specific biases in FUS2M are small enough, then the practical bottleneck for medical foundation models shifts from pretraining methodology to data curation and governance, which may change where funding and effort are best spent.
  • A direct test of the paper's interpretation would be to pretrain DINOv2 on a non-Danish fetal ultrasound corpus and compare cross-country transfer; the paper does not perform this ablation, so its generalization claim remains to be stress-tested.
  • The observation that attention heads specialize to anatomy suggests that domain-specific self-supervised learning may produce features that are useful beyond the three benchmark tasks, for example in anomaly detection or quality control of screening images, though those uses are not evaluated here.
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

3 major / 5 minor

Summary. The paper addresses a practical design question for medical foundation models: whether to adapt an existing natural-image foundation model, and whether custom pretraining requires novel self-supervised methods. The authors pretrain ViT-S and ViT-B backbones with DINOv2-style self-distillation from scratch on FUS2M, a 2M-image Danish fetal ultrasound screening dataset, creating UltraDINO. They evaluate on standard-plane classification (Fetal Planes, Spain), abdominal structure segmentation (FASS, Brazil; full, 20-image few-shot, and label-efficiency settings), and few-shot pubic symphysis/fetal head segmentation (JNU-IFM, China), comparing with DINOv2 and iBOT checkpoints, the USFM ultrasound foundation model, and from-scratch baselines. The headline results are large gains for UltraDINO in few-shot and linear-probing settings and smaller gains in fully finetuned tasks. The paper concludes that domain-specific pretraining is worth the compute and that mature CV methods should be preferred over novel methodological development.

Significance. The empirical setup is strong in several respects: the pretraining data is large and clinically sourced, the downstream evaluation is external and multicountry, and the comparison set includes natural-image SSL models of different scales, a published ultrasound foundation model, and supervised baselines. The evaluation is not circular, since FUS2M is not used for finetuning and all downstream datasets are public and non-overlapping with the pretraining data. If the results stand, the practical recommendation—use a proven SSL recipe on a large, domain-specific pretraining set rather than designing new pretraining machinery—would be valuable for medical imaging groups with limited compute. The released code improves reproducibility. However, the headline methodological claim ('avoid novel methods') is currently supported only by a confounded comparison, and several reported advantages are small relative to the missing variance information.

major comments (3)
  1. [§3.1, Table 1, Limitations, §5] The second key insight and the prescriptive conclusion rest on a method/data confound. UltraDINO is DINOv2 self-distillation trained from scratch on 2M fetal ultrasound images, while the only domain-specific novel-method baseline, USFM, is evaluated using published weights trained on a different ~2M multi-organ ultrasound dataset; the authors state in §3.1 that 'we were unable to reproduce the method on FUS2M.' The large gaps in Table 1 (e.g., linear-probing F1 95.06 vs 42.78; JNU-IFM Dice 86.01 vs 83.05) are therefore equally compatible with the advantage coming from fetal-specific, high-quality pretraining data rather than from the choice of DINOv2 over a novel ultrasound-specific method. To support the conclusion that mature methods are sufficient and novel methods should be avoided, the manuscript needs a same-data comparison: either train the USFM-style (or another novel SSL) method on FUS2M, or train the DINOv2 recipe on USFM's pretraining corpus, or explicitly weaken the abstract and conclusion to a data-value claim rather than a method-value claim.
  2. [Table 1, §3.1] Several differences behind the 'state-of-the-art' and 'consistently outperforms' claims are smaller than the noise one would expect from three seeds and five folds, yet no standard deviations, confidence intervals, or significance tests are reported. For example, on FASS full-image segmentation UltraDINO ViT-B scores 79.51 Dice versus USFM 78.86 and DINOv2 79.26; on Fetal Planes fine-tuning the F1 gap is 94.54 versus 94.16/93.94. Without variance estimates the reader cannot assess whether these differences are meaningful, and even the few-shot gaps (e.g., JNU-IFM 86.01 vs 83.55) should be reported with their spread. Please report per-seed/per-split results or error bars and state whether the observed gaps are statistically distinguishable.
  3. [§4, Generalizability claim] The claim that 'the pretraining dataset is not regionally overlapping with the finetuning tasks, indicating strong generalization capability' is asserted without quantitative support. FUS2M is a Danish regional screening dataset acquired on GE Logiq/Voluson scanners, whereas the evaluation datasets come from different countries and likely different scanners and protocols; the manuscript reports no scanner information for the evaluation sets and no analysis of distribution shift. If the authors wish to claim strong cross-country generalization, they should either report acquisition details and any domain-shift statistics, or soften the claim to 'evaluation on non-overlapping public datasets'.
minor comments (5)
  1. [§2, §4, Abstract] The phrase 'requiring no hyperparameter tuning' should be qualified: the authors did not tune hyperparameters, but DINOv2 itself was tuned extensively on natural images; the finding is that the default recipe transferred to ultrasound, not that the method requires no tuning in general.
  2. [Table 1, §3.1] DINOv2 checkpoints use patch size 14 while all other models use patch size 16, which changes the effective token count at 224×224 resolution; the potential effect on comparability should be discussed or controlled.
  3. [§4, Figure 4] The attention-map comparison is anecdotal; consider quantifying head specialization (e.g., with a diversity or localization metric) or clearly marking the figure as illustrative only.
  4. [§3.3] One typo: 'validaiton' should be 'validation'.
  5. [§1, Code availability] The paper states that code is available but does not explicitly state whether the UltraDINO pretrained weights are released; please specify, since weight release is central to the reproducibility of a foundation-model case study.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: UltraDINO's pretraining and evaluation are cleanly separated; no fitted input is renamed as a prediction.

full rationale

The paper conducts an empirical case study: UltraDINO is trained from scratch with DINOv2 self-distillation on FUS2M (2M Danish fetal ultrasound images) and evaluated on external public datasets (FASS from Brazil, JNU-IFM from China, Fetal Planes from Spain). None of these evaluation sets are part of pretraining, and no parameter of UltraDINO is fitted to the evaluation data. Comparisons use published checkpoints for DINOv2, iBOT, and USFM, i.e., external independent artifacts, not a self-citation chain. The two key insights (domain-specific pretraining helps; mature CV methods suffice) are empirical comparisons, not derivations that reduce to their inputs. The acknowledged USFM limitation (no public code, so it could not be retrained on FUS2M) creates a data/method confound in the comparison to USFM; however, that is a fairness or generalizability concern about experimental design, not a circularity in which an output equals an input by construction. No self-definitional equation, no fitted-input-called-prediction, and no uniqueness theorem imported from the authors' own prior work appear in the manuscript. The paper is self-contained against external benchmarks, so the appropriate finding is no significant circularity.

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

The central claim rests on domain assumptions about data representativeness, method transferability, and annotation quality, rather than on fitted parameters or invented entities.

assumptions (4)
  • domain assumption FUS2M is a high-quality, representative sample of fetal ultrasound screening images.
    The paper uses FUS2M as the pretraining corpus and claims it represents the fetal ultrasound domain; no explicit quality filtering beyond machine types and exclusion of ~10k images from prior studies (Section 2).
  • domain assumption The DINOv2 self-distillation recipe produces useful features in fetal ultrasound when trained from scratch.
    The paper's central result depends on this method transfer; they did not tune hyperparameters and used the default configuration (Section 2, Configuration).
  • domain assumption Ground-truth annotations in FetalPlanes, FASS, and JNU-IFM are accurate and consistent across countries.
    Evaluation relies on these public datasets; annotation quality is taken as given (Section 3).
  • domain assumption USFM's released weights are a fair representative of the ultrasound foundation model class in this comparison.
    USFM was not retrained on FUS2M because its code is unavailable; comparison uses the authors' published weights (Section 3.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of General Methods Make Great Domain-specific Foundation Models: A Case-study on Fetal Ultrasound." pith.science (2026). https://pith.science/paper/ZJSO4KA2

@misc{pith2026250619552,
  author       = {Pith},
  title        = {Pith review of: General Methods Make Great Domain-specific Foundation Models: A Case-study on Fetal Ultrasound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZJSO4KA2}},
  note         = {Machine review of arXiv:2506.19552}
}
read the original abstract

With access to large-scale, unlabeled medical datasets, researchers are confronted with two questions: Should they attempt to pretrain a custom foundation model on this medical data, or use transfer-learning from an existing generalist model? And, if a custom model is pretrained, are novel methods required? In this paper we explore these questions by conducting a case-study, in which we train a foundation model on a large regional fetal ultrasound dataset of 2M images. By selecting the well-established DINOv2 method for pretraining, we achieve state-of-the-art results on three fetal ultrasound datasets, covering data from different countries, classification, segmentation, and few-shot tasks. We compare against a series of models pretrained on natural images, ultrasound images, and supervised baselines. Our results demonstrate two key insights: (i) Pretraining on custom data is worth it, even if smaller models are trained on less data, as scaling in natural image pretraining does not translate to ultrasound performance. (ii) Well-tuned methods from computer vision are making it feasible to train custom foundation models for a given medical domain, requiring no hyperparameter tuning and little methodological adaptation. Given these findings, we argue that a bias towards methodological innovation should be avoided when developing domain specific foundation models under common computational resource constraints.

Figures

Figures reproduced from arXiv: 2506.19552 by the authors.

Figure 1
Figure 1. Segmentation performance depending on model and pretraining dataset. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. UltraDINO uses the self-distillation method used by DINOv2, image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. iBOT-loss curves for two runs with the same configurations [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Attention maps of DINOv2 and UltraDINO ViT-B models: UltraDINO [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. DIET-CP: Lightweight and Data Efficient Self Supervised Continued Pretraining

    cs.CV 2025-09 conditional novelty 6.0 of 10

    DIET-CP continues pretraining a vision model by classifying each image by its index on roughly 1000 unlabeled images, improving medical and astronomical classification while degrading fine-grained natural-image classi...

Reference graph

Works this paper leans on

27 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    arXiv preprint arXiv:2108.07258 (2021)

    Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., et al.: On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258 (2021)

    arXiv 2021

  2. [2]

    Scientific Reports10(1), 10200 (2020)

    Burgos-Artizzu, X.P., Coronado-Gutiérrez, D., Valenzuela-Alcaraz, B., Bonet- Carne, E., Eixarch, E., Crispi, F., Gratacós, E.: Evaluation of deep convolutional neural networks for automatic classification of common maternal fetal ultrasound planes. Scientific Reports10(1), 10200 (2020)

    work page 2020

  3. [3]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  4. [4]

    Correggio, K.S.D., Galluzzo, R.N., Santos, L.O., Barroso, F.S.M., Chaves, T.Z.L., Onofre, A.S.C., von Wangenheim, A.: Fetal abdominal structures segmenta- tion dataset using ultrasonic images (Oct 2023).https://doi.org/10.17632/ 4gcpm9dsc3.1,https://data.mendeley.com/datasets/4gcpm9dsc3/1

    work page 2023

  5. [5]

    arXiv preprint arXiv:2010.11929 (2020) 10 Ambsdorf et al

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 10 Ambsdorf et al

    arXiv 2020

  6. [6]

    Advances in neural information processing systems33, 21271–21284 (2020)

    Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Do- ersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems33, 21271–21284 (2020)

    2020

  7. [7]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: Proceedings of the IEEE/CVF winter conference on applications of computer vi- sion. pp. 574–584 (2022)

    2022

  8. [8]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000–16009 (2022)

    2022

Show all 27 references
  1. [9]

    Huix, J.P., Ganeshan, A.R., Haslum, J.F., Söderberg, M., Matsoukas, C., Smith, K.: Are natural domain foundation models useful for medical image classification? In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 7634–7643 (2024)

    2024

  2. [10]

    Neural networks121, 74–87 (2020)

    Ibtehaz, N., Rahman, M.S.: Multiresunet: Rethinking the u-net architecture for multimodal biomedical image segmentation. Neural networks121, 74–87 (2020)

    2020

  3. [11]

    Nature methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

    2021

  4. [12]

    arXiv preprint arXiv:2411.16380 (2024)

    Jiang, Y., Feng, C.M., Ren, J., Wei, J., Zhang, Z., Hu, Y., Liu, Y., Sun, R., Tang, X., Du, J., et al.: Privacy-preserving federated foundation model for generalist ultrasound artificial intelligence. arXiv preprint arXiv:2411.16380 (2024)

    2024

  5. [13]

    Medical Image Analysis 96, 103202 (2024)

    Jiao,J.,Zhou,J.,Li,X.,Xia,M.,Huang,Y.,Huang,L.,Wang,N.,Zhang,X.,Zhou, S., Wang, Y., et al.: Usfm: A universal ultrasound foundation model generalized to tasks and organs towards label efficient image analysis. Medical Image Analysis 96, 103202 (2024)

    2024

  6. [14]

    In: International Workshop on Applications of Medical AI

    Juodelyte, D., Lu, Y., Jiménez-Sánchez, A., Bottazzi, S., Ferrante, E., Cheplygina, V.: Source matters: Source dataset impact on model robustness in medical imaging. In: International Workshop on Applications of Medical AI. pp. 105–115. Springer (2024)

    2024

  7. [15]

    Kang, Q., Lao, Q., Gao, J., Bao, W., Lu, Q., Li, K.: Urfm: a general ultrasound representation foundation model for advancing ultrasound image diagnosis (2024)

    2024

  8. [16]

    Nature Biomedical Engineering6(12), 1346–1352 (2022)

    Krishnan, R., Rajpurkar, P., Topol, E.J.: Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering6(12), 1346–1352 (2022)

    2022

  9. [17]

    Data in brief41, 107904 (2022)

    Lu, Y., Zhou, M., Zhi, D., Zhou, M., Jiang, X., Qiu, R., Ou, Z., Wang, H., Qiu, D., Zhong, M., et al.: The jnu-ifm dataset for segmenting pubic symphysis-fetal head. Data in brief41, 107904 (2022)

    2022

  10. [18]

    Na- ture616(7956), 259–265 (2023)

    Moor, M., Banerjee, O., Abad, Z.S.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Na- ture616(7956), 259–265 (2023)

    2023

  11. [19]

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res. (2024)

    2024

  12. [20]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T., Efros, A.A.: Context en- coders: Feature learning by inpainting. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2536–2544 (2016)

    2016

  13. [21]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

    2015

  14. [22]

    arXiv preprint arXiv:2402.15687 (2024)

    Song, X., Xu, X., Yan, P.: General purpose image encoder dinov2 for medical image registration. arXiv preprint arXiv:2402.15687 (2024)

    2024 arXiv

  15. [23]

    arXiv preprint arXiv:2410.23132 (2024)

    Wald, T., Ulrich, C., Lukyanenko, S., Goncharov, A., Paderno, A., Maerkisch, L., Jäger, P.F., Maier-Hein, K.: Revisiting mae pre-training for 3d medical image segmentation. arXiv preprint arXiv:2410.23132 (2024)

    2024 arXiv

  16. [24]

    arXiv preprint arXiv:2308.02463 (2023)

    Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Towards generalist foundation model for radiology by leveraging web-scale 2d&3d medical data. arXiv preprint arXiv:2308.02463 (2023)

    2023 arXiv

  17. [25]

    In: European Conference on Computer Vision

    Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: European Conference on Computer Vision. Springer (2018)

    2018

  18. [26]

    arXiv preprint arXiv:2303.00915 (2023)

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023)

    2023 arXiv

  19. [27]

    arXiv preprint arXiv:2111.07832 (2021)

    Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832 (2021)

    2021 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.