Pith. sign in

REVIEW 6 major objections 8 minor 38 references

Can Foundation Models Generalise the Presentation Attack Detection Capabilities on ID Cards?

T0 review · 6 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper claims that representative bona fide images, not large attack sets, are the key to generalising presentation attack detection on ID cards, and that DinoV2 features with a one-layer head deliver the best results.

desk verdict Useful benchmark of DinoV2/CLIP for ID-card PAD, but the headline claim about bona fide images is not supported by the experiments as designed. read the letter →

arxiv 2506.05263 v1 pith:TMJ3S67M submitted 2025-06-05 cs.CV

classification cs.CV
keywords presentationattackdetectionIDcardfraudfoundationmodelsDinoV2generalisationbonafideimageszero-shotlearningscore-levelfusion
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 claims that the key to making presentation attack detection (PAD) for ID cards generalise to unseen countries is having representative bona fide (genuine) images, not a huge number of attack samples. It benchmarks foundation models—DinoV2 and CLIP—as frozen feature extractors with a single-layer fine-tuned head, under zero-shot, leave-one-out, and two-class protocols on two datasets: a private Chilean ID dataset (CHL-DBP) and the open-set ID-Net (Finland, Spain, Slovakia). DinoV2 consistently outperforms CLIP and conventional deep models; on CHL-DBP it reaches 6.33% EER, and fusing DinoV2 features at score level with DenseNet improves ID-Net from 15.65% to 8.25% EER. The authors attribute ID-Net's higher error to its synthetic bona fide images, which resemble the synthetic attacks and confuse the classifier. If correct, the practical implication is that PAD generalisation is primarily a data-acquisition problem about genuine ID captures, which is actionable for industry.

What carries the argument

The load-bearing machinery is DinoV2 (a vision transformer trained self-supervised on a large image corpus) used as a frozen feature extractor, with a classification head of one fully-connected layer plus sigmoid trained on the PAD binary task; the head is the only part fine-tuned, keeping the foundation model's weights fixed. This is complemented by a score-level fusion that averages DinoV2's output score with that of DenseNet121, combining local patch-level traces (DinoV2) with global document structure (DenseNet). The evaluation protocols—zero-shot, leave-one-out for unknown attack types, and two-class fine-tuning—are the instruments that expose generalisation behaviour across datasets.

What would settle it

Take the CHL-DBP attack images (border, printed, screen) and replace only the bona fide training images with synthetically generated ID templates of the type used in ID-Net, keeping all other protocol details fixed; if the error rate does not rise toward the 27% level, then genuine-image representativeness is not the cause of the gap.

Watch

Extended reading notes

Core claim

The paper's central claim is that a well-chosen foundation model, used as a feature extractor and fine-tuned with a minimal one-layer head, generalises PAD on ID cards across countries at least as well as purpose-built deep networks, and that the decisive factor is the quality and representativeness of bona fide images rather than the quantity or diversity of attack types. The strongest evidence is DinoV2-large reaching 6.33% EER on the private CHL-DBP dataset and 27.86% on ID-Net, versus the best conventional networks (ResNet101: 31.05% on ID-Net; SwinT: 33.27%). Because ID-Net's genuine images are synthetically generated from templates with Stable Diffusion while its attacks are also digitally generated, the authors conclude that synthetic bona fide data confuses the model; adding a score-level fusion with DenseNet121, which captures more global cues, reduces ID-Net EER to 8.25%. The paper frames the task as catching local traces and artefacts on real ID documents, which requires genuine captures, not synthetic ones.

Load-bearing premise

The claim that representative bona fide images are the key to generalisation rests on comparing two datasets that differ in attack types, document layouts, and capture conditions at the same time, so the observed performance gap is not isolated to the bona fide images.

Editorial extensions

If this is right

  • Deployment teams can expect that collecting genuine ID captures from each new country matters more than amassing attack samples from many countries.
  • A foundation model frozen as feature extractor plus a one-layer head is a cheap adaptation path when privacy rules limit access to real ID data.
  • Score-level fusion of local and global feature sources is a practical way to close part of the gap on open-set, synthetically generated ID datasets.
  • Synthetic bona fide images, as in ID-Net, should be treated with caution for training; they may be better suited as test data.
  • The leave-one-out protocol gives a realistic estimate of performance on unseen attack types, which is the operational scenario for remote onboarding.

Reading between the lines

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

  • If the bona fide representativeness claim is right, then data-generation efforts should pivot from synthesising attacks to synthesising very realistic genuine ID cards, evaluated by how well they close the gap on real-capture datasets like CHL-DBP.
  • The 8.25% fusion result on ID-Net suggests that global document cues (layout, background, printing artefacts) and local patch cues are complementary; a single model that jointly models both scales might outperform either branch alone.
  • The authors' explanation for ID-Net's higher error is confounded: ID-Net also differs in attack types (digital tampering vs physical recapture) and document variety (three countries vs one), so a controlled test that swaps only bona fide source would settle the attribution.
  • Since CLIP lags DinoV2 despite similar scale, the self-supervised local feature learning in DinoV2 (patch tokens, no language alignment) appears better suited to forgery trace detection; text-prompt CLIP variants remain an unexplored avenue.
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

6 major / 8 minor

Summary. The paper benchmarks foundation models (DinoV2, CLIP) and classical deep networks for presentation attack detection (PAD) on ID cards using two datasets: a private Chilean ID-card dataset (CHL-DBP) with physical attacks (border, printed, screen) and the synthetic ID-Net dataset with digital tampering attacks (combined, face-swap, copy-move, face-morphing). It evaluates zero-shot, fine-tuned, and leave-one-out-attack protocols, and proposes a score-level fusion of DinoV2 and DenseNet to improve results on ID-Net. The headline claim, stated in the introduction and abstract, is that representative bona fide images are the key to generalisation across countries.

Significance. If the headline claim were established, the paper would provide actionable guidance for PAD data collection and a useful benchmark for foundation models on ID-card documents. The extensive comparisons under zero-shot and fine-tuning on two datasets are potentially valuable, and the LOO protocol for unknown attacks is a reasonable design. The consistent advantage of DinoV2 over CLIP and classical models is an interesting empirical result. However, the causal claim about bona fide images is not tested, fusion selection is performed on the test set, and no confidence intervals are reported, so the quantitative conclusions must be treated with caution.

major comments (6)
  1. [I, VII, Table VI] The claim in Section I that 'the key for generalisation capabilities is to have representative bona fide images' is not supported. Table VI compares CHL-DBP (real bona fide, physical attacks) with ID-Net (synthetic bona fide, digital tampering attacks), so the datasets differ simultaneously in bona fide realism, attack types, capture conditions, and document layouts. Section VII itself notes that ID-Net's attacks are generated with techniques similar to its bona fide images, which is an alternative explanation for the higher EER. No experiment varies bona fide fidelity while holding attacks fixed, so the causal attribution is untested; this directly affects the abstract and contribution list.
  2. [VI-B, VI-C] The leave-one-out protocol holds out attack types (border, printed, screen on CHL-DBP; face-replacement, combined, face-morphing, copy-move on ID-Net), not countries. The stated motivation in Section I concerns generalisation to an 'unknown new ID card country,' and the title asks about generalisation of PAD capabilities on ID cards. No experiment trains on some countries and tests on a held-out country, even though ID-Net contains three countries; the country-level claim is therefore not directly tested.
  3. [VI-E, Table VII] The fusion system (DinoV2-vitb14 + DenseNet121) was selected after evaluating individual models on the ID-Net test set, and its 8.25% EER is reported on that same test set. Without a held-out selection set or nested cross-validation, the fusion result is optimistically biased. Please describe the model-selection protocol explicitly and report results on an independent test set or with cross-validation.
  4. [VI-A] The text describes classifiers 'trained from scratch' but the same sentence says they are initialized with ImageNet weights. This contradicts the definition of a from-scratch baseline. Since all models in the comparison use ImageNet pretraining, the claimed advantage of foundation models over training from scratch is not measured; rename the baseline as ImageNet-initialized fine-tuning or train truly from scratch.
  5. [Table IV] Table IV reports DinoV2-vitl14 with 23.73% EER for the Border LOO fold, whereas the same model achieves 4.33% in the zero-shot setting (Table III) and 6.33% in the two-class fine-tuning setting (Table VI). This value is either a transcription error or needs an explanation, since as printed it contradicts the general pattern that larger DinoV2 models perform better.
  6. [Tables III-VII] No confidence intervals, repeated runs, or statistical significance tests are reported for any of the tables. Given the variability typical of biometric PAD evaluations, the ranking of models might change; please provide error bars or significance tests for the main comparisons.
minor comments (8)
  1. [Table V] Table V's header repeats the CHL-DBP attack names ('BORDER, PRINTED AND SCREEN') for the ID-Net experiment; it should list the ID-Net attack types (Face-replacement, Face combined, Face-morphing, Face Copy-Move).
  2. [VI-D] Section VI-D first says the ID-Net two-class evaluation uses 'border printed and screen attack,' which do not exist in ID-Net; the following sentence correctly lists combined, copy-move, face-morphing, and face-replacement, but the inconsistency should be fixed.
  3. [VI-E] The reference to 'Figure VII' is unclear; the manuscript contains Figures 1-5, so the sentence should point to the correct figure or the missing figure should be added.
  4. [VI-C] The model name 'Dinov2-visl14' appears to be a typo for DinoV2-vitl14.
  5. [VI-B] 'MetaClip-400-m' is not a standard name; CLIP was trained on 400 million image-text pairs, so the reference should be corrected.
  6. [Table II] The 'Total per set' row for CHL-DBP lists 8,173, 24,602, 8,172, which are actually the Test, Train, and Validation column totals, respectively; the row is mislabeled and should be corrected.
  7. [Table III] The dataset name is abbreviated as 'CHL-DB' in Table III; it should be 'CHL-DBP'.
  8. [Figure 4] 'Leave-out-protocol' should be 'Leave-One-Out protocol.'

Circularity Check

1 steps flagged · score 2.0 of 10

The fusion EER is selected on the same test set it reports, but the main generalisation claim is confounded rather than circular.

  1. fitted input called prediction [Section VI-E (Fusion level) and Table VII]
    "To enhance the ID-Net results, we suggest integrating the method that achieved the lowest EER from the best-trained model developed from scratch based on deep learning and Imagenet weights, which detects more global features, with the top-performing foundation model based on DinoV2, which focuses on local features, as illustrated in Figure 3. ... The results previously obtained were reduced from 15.65% EER obtained for ID-Net to a new EER of 8.25%."

    The two fusion partners are selected as the 'lowest EER' deep model and the 'top-performing' DinoV2 variant using the same test-set EER tables (Tables V and VI) that are then used to report the fusion's 8.25% EER on ID-Net. The model choice is therefore a hyperparameter selected on the test labels, and the reported number is the value of that selection criterion on the same set, not an independent evaluation. The fusion EER is thus optimistically biased by construction and cannot serve as a predictive claim that the fusion 'boosts' ID-Net results.

full rationale

The paper is an empirical benchmark rather than a derivation, so most of its claims are not circular: the LOO EERs are measured outputs, DinoV2 and CLIP are external pre-trained models, and no equation is defined in terms of the result it is supposed to establish. The main circularity-adjacent issue is the fusion experiment: DinoV2-vitb14 is chosen as the 'top-performing' model and DenseNet121 as the 'lowest EER' deep model using the same test-set tables that are then used to report the 8.25% EER, so that number is a post-selection estimate on the same labels, not an independent prediction. This is a test-set-selection bias in one secondary result. The paper's headline conclusion, that representative bona fide images are the key to generalisation, is not circular, but it is confounded: the CHL-DBP versus ID-Net comparison varies attack types, capture conditions, and document layout together with bona fide realism, so the causal attribution is untested. That is a validity concern, not a circularity, and does not raise the score above 2.

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

This is an empirical benchmark, so the ledger contains no derivational constants. The central claim rests on a two-dataset comparison, a LOO protocol, and threshold-based metrics. Hand-set constants are: grid-searched hyperparameters (lr 1e-3, batch 64, 100 epochs, one-layer head), per-protocol decision thresholds fitted to evaluation data, and the equal-weight score average in the fusion. The headline 'bona fide is key' conclusion additionally assumes, without a controlled experiment, that the CHL-DBP vs ID-Net gap is attributable to bona fide image representativeness. No invented entities.

free parameters (3)
  • Decision threshold per protocol. = Not reported (EER operating point).
    EER and BPCER10/20/100 in Tables III-VII are computed at thresholds fitted to the same evaluation data; no fixed operating point or confidence intervals are given (Section V).
  • Fusion weight (score average). = 0.5 DinoV2 + 0.5 DenseNet.
    Section VI-E fuses by simple average chosen by hand after inspecting test-set results; not validated on held-out data.
  • Fine-tuning hyperparameters. = lr 1e-3, batch 64, 100 epochs, one-layer head.
    Selected by grid search on validation (Section VI-C); reported only for the winning configuration, so the gain over other configurations is not quantified.
assumptions (4)
  • domain assumption The leave-one-out protocol, training on two attack types and testing on the third, validly estimates generalization to unknown attacks.
    Section VI defines LOO for CHL-DBP and applies it to ID-Net. The generalization claim assumes the held-out attack is representative of unseen attacks and that training on the other two types does not leak into the test split.
  • domain assumption ISO/IEC 30107-3 APCER/BPCER with thresholds fitted to the evaluation set is the correct performance measure for operational PAD.
    Section V defines the metrics. Reported EER and BPCER10/20/100 are computed at data-dependent thresholds, so interpreting them as operational performance assumes the fitted operating point transfers to deployment.
  • ad hoc to paper The CHL-DBP versus ID-Net performance gap is caused by bona fide image representativeness (real vs synthetic), not by attack type, capture condition, or document layout differences.
    Stated in Section VII ('The private dataset ID-NET uses as a genuine ID card image an image created from an ID template modified by stable diffusion models') and Section VI-D. No experiment holds attacks fixed while varying bona fide quality, so the attribution is assumed rather than measured.
  • ad hoc to paper ImageNet-initialized networks that are then fine-tuned constitute the 'trained from scratch' baseline.
    Section VI-A says classifiers were 'trained from scratch based on ImageNet weights', which is contradictory; the baseline is fine-tuning from ImageNet, so conclusions about FM superiority versus true from-scratch training do not follow.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Foundation Models Generalise the Presentation Attack Detection Capabilities on ID Cards?." pith.science (2026). https://pith.science/paper/TMJ3S67M

@misc{pith2026250605263,
  author       = {Pith},
  title        = {Pith review of: Can Foundation Models Generalise the Presentation Attack Detection Capabilities on ID Cards?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TMJ3S67M}},
  note         = {Machine review of arXiv:2506.05263}
}
read the original abstract

Nowadays, one of the main challenges in presentation attack detection (PAD) on ID cards is obtaining generalisation capabilities for a diversity of countries that are issuing ID cards. Most PAD systems are trained on one, two, or three ID documents because of privacy protection concerns. As a result, they do not obtain competitive results for commercial purposes when tested in an unknown new ID card country. In this scenario, Foundation Models (FM) trained on huge datasets can help to improve generalisation capabilities. This work intends to improve and benchmark the capabilities of FM and how to use them to adapt the generalisation on PAD of ID Documents. Different test protocols were used, considering zero-shot and fine-tuning and two different ID card datasets. One private dataset based on Chilean IDs and one open-set based on three ID countries: Finland, Spain, and Slovakia. Our findings indicate that bona fide images are the key to generalisation.

Figures

Figures reproduced from arXiv: 2506.05263 by the authors.

Figure 1
Figure 1. Top: Example of images of the PAD Chilean ID cards CHL-DBP dataset. Left to right: Bona fide, Composite, Printed and Screen. Bottom: ID-Net example of open-set dataset (ID-Net). Left to right: Bona fide, combined, face-swap, copy-move attack [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. ID card Presentation Attack Detection pipeline. The ViT network divides the image into patches. The embedding space of the FM is adapted by [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. ID card Presentation Attack Detection using a fusion pipeline. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: From left to right: images patch sample zoomed for bona fide, printed [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: Summary of the DET curves for fine-tuning and Leave-out-protocol for the two datasets considering Deep learning, Vision-Transformer and Foundation [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages

  1. [1]

    Few-shot object detection: A comprehensive survey,

    M. K ¨ohler, M. Eisenbach, and H.-M. Gross, “Few-shot object detection: A comprehensive survey,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 9, pp. 11 958–11 978, 2024

  2. [2]

    SynID: Passport synthetic dataset for presentation attack detection,

    J. E. Tapia, F. Stockhardt, L. J. Gonz ´alez-Soler, and C. Busch, “SynID: Passport synthetic dataset for presentation attack detection,” 2025. [Online]. Available: https://arxiv.org/abs/2505.07540

  3. [3]

    Open source face image quality (OFIQ) - implementation and evaluation of algorithms,

    J. Merkle, C. Rathgeb, B. Herdeanu, B. Tams, D. Lou, A. D ¨orsch, M. Schaubert, J. Dehen, L. Chen, X. Yin, D. Huang, A. Strat- mann, M. Ginzler, M. Grimmer, and C. Busch, “Open source face image quality (OFIQ) - implementation and evaluation of algorithms,” https://www.bsi.bund.de/SharedDocs/Downloads/EN/BSI/ OFIQ/Projektabschlussbericht OFIQ 1 0.pdf, Sep...

  4. [4]

    Face analysis technology evaluation (FATE) part 11: Face image quality vector assessment - specific image defect detection,

    J. Yang, P. Grother, M. Ngan, K. Hanaoka, and A. Hom, “Face analysis technology evaluation (FATE) part 11: Face image quality vector assessment - specific image defect detection,” National Institute of Standards and Technology, NIST Interagency Report 8485, 2024

  5. [5]

    Forged presentation attack detection for ID cards on remote verification systems,

    S. Gonzalez and J. E. Tapia, “Forged presentation attack detection for ID cards on remote verification systems,” Pattern Recognition , vol. 162, p. 111352, 2025. [Online]. Available: https://www.sciencedirect. com/science/article/pii/S0031320325000123

  6. [6]

    MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream,

    V . Arlazarov, K. Bulatov, T. Chernov, and V . Arlazarov, “MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream,” Computer Optics , vol. 43, pp. 818–824, 10 2019

  7. [7]

    MIDV-2020: a comprehensive benchmark dataset for identity document analysis,

    B. Konstantin, E. E, T. Daniil, S. Natalya, C. Yulia, S. Alexander, U. S, M. Zuheng, B. Jean-Christophe, L. Muzzamil, and A. Vladimir, “MIDV-2020: a comprehensive benchmark dataset for identity document analysis,” Computer Optics , vol. 46, pp. 252–270, 03 2022

  8. [8]

    Document Liveness Challenge DLC-2021 - part 1 (or, cg),

    E. D. Polevoy Dimitry and, Sigareva Irina and, A. Vladimir, N. Dmitry, M. Zuheng, L. Muhammad, and B. Jean-Christophe, “Document Liveness Challenge DLC-2021 - part 1 (or, cg),” May 2022. [Online]. Available: https://doi.org/10.5281/zenodo.6586764

Show all 38 references
  1. [9]

    MIDV-holo: A dataset for id document hologram detection in a video stream,

    L. Koliaskina and E. et al., “MIDV-holo: A dataset for id document hologram detection in a video stream,” in International Conference on Document Analysis and Recognition . Springer, 2023, pp. 486–503

  2. [10]

    IDNet: A novel dataset for identity document analysis and fraud detection (part 2),

    H. Guan, Y . Wang, L. Xie, S. Nag, R. Goel, N. E. N. Swamy, Y . Yang, C. Xiao, J. Prisby, R. Maciejewski, and J. Zou, “IDNet: A novel dataset for identity document analysis and fraud detection (part 2),” Feb. 2024. [Online]. Available: https://doi.org/10.5281/zenodo.10602369

  3. [11]

    Hybrid two-stage architecture for tampering detection of chipless ID cards,

    S. Gonzalez, A. Valenzuela, and J. Tapia, “Hybrid two-stage architecture for tampering detection of chipless ID cards,” IEEE Transactions on Biometrics, Behavior , and Identity Science , vol. 3, no. 1, pp. 89–100, 2021

  4. [12]

    Synthetic ID card image generation for improving presentation attack detection,

    D. Benalcazar, J. Tapia, G. Sebastian, and C. Busch, “Synthetic ID card image generation for improving presentation attack detection,” IEEE Transactions on Information F orensics and Security , vol. 18, pp. 1814– 1824, 2023

  5. [13]

    Open-set: ID card presentation attack detection using neural style transfer,

    R. P. Markham, J. M. E. L ´opez, M. Nieto-Hidalgo, and J. E. Tapia, “Open-set: ID card presentation attack detection using neural style transfer,” IEEE Access , vol. 12, pp. 68 573–68 585, 2024

  6. [14]

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

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in Computer Vision (ICCV), 2017 IEEE International Conference on , 2017

  7. [15]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.-Y . Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on , 2017

  8. [16]

    Few-shot learning: Expanding ID cards presentation attack detection to unknown ID countries,

    A. Sanchez, J. M. Esp ´ın, and J. E. Tapia, “Few-shot learning: Expanding ID cards presentation attack detection to unknown ID countries,” in 2024 IEEE International Joint Conference on Biometrics (IJCB) , 2024, pp. 1– 9

  9. [17]

    First competition on presentation attack detection on ID card,

    J. E. Tapia, N. Damer, C. Busch, J. M. Espin, J. Barrachina, A. S. Rocamora, K. Ocvirk, L. Alessio, B. Batagelj, S. Patwardhan, R. Ra- machandra, R. Mudgalgundurao, K. Raja, D. Schulz, and C. Aravena, “First competition on presentation attack detection on ID card,” in 2024 IEE...

  10. [18]

    Domain-agnostic document authentication against practical recapturing attacks,

    C. Chen, S. Zhang, F. Lan, and J. Huang, “Domain-agnostic document authentication against practical recapturing attacks,” IEEE Transactions on Information F orensics and Security , vol. 17, pp. 2890–2905, 2022

  11. [19]

    Moire spectral augmentation and masked ´ frequency modeling for document presentation attack detection,

    C. Chen, Y . Li, B. Li, W. Yu, B. Chen, B. Li, and J. Huang, “Moire spectral augmentation and masked ´ frequency modeling for document presentation attack detection,” IEEE Transactions on Dependable and Secure Computing, pp. 1–16, 2025

  12. [20]

    KID34K: A dataset for online identity card fraud detection,

    E.-J. Park, S.-Y . Back, J. Kim, and S. S. Woo, “KID34K: A dataset for online identity card fraud detection,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, ser. CIKM ’23. New York, NY , USA: Association for Computing Machinery...

  13. [21]

    MIDV-2019: Challenges of the modern mobile-based document ocr,

    K. Bulatov, D. Matalov, and V . Arlazarov, “MIDV-2019: Challenges of the modern mobile-based document ocr,” 10 2019

  14. [22]

    Synthetic dataset of ID and travel document,

    C. Boned, M. Talarmain, N. Ghanmi, G. Chiron, S. Biswas, A. M. Awal, and O. R. Terrades, “Synthetic dataset of ID and travel document,” 2024. [Online]. Available: https://arxiv.org/abs/2401.01858

  15. [23]

    Guilloche detection for ID authentication: A dataset and baselines,

    M. Al-Ghadi, Z. Ming, P. Gomez-Kr ¨amer, J.-C. Burie, M. Coustaty, and N. Sidere, “Guilloche detection for ID authentication: A dataset and baselines,” in 2023 IEEE 25th International Workshop on Multimedia Signal Processing (MMSP) , 2023, pp. 1–6

  16. [24]

    Foundation models and biometrics: A survey and outlook,

    H. O. Shahreza and S. Marcel, “Foundation models and biometrics: A survey and outlook,” TechRxiv, 2025

  17. [25]

    Froundation: Are foundation models ready for face recognition?

    T. Chettaoui, N. Damer, and F. Boutros, “Froundation: Are foundation models ready for face recognition?” 2024. [Online]. Available: https://arxiv.org/abs/2410.23831

  18. [26]

    Foundpad: Foundation models reloaded for face presen- tation attack detection,

    G. Ozgur, E. Caldeira, T. Chettaoui, F. Boutros, R. Ramachandra, and N. Damer, “Foundpad: Foundation models reloaded for face presen- tation attack detection,” in Proceedings of the Winter Conference on Applications of Computer Vision (WACV) Workshops , February 2025, pp. 745–755

  19. [27]

    Towards iris presentation attack detection with foundation models,

    J. E. Tapia, L. J. Gonz ´alez-Soler, and C. Busch, “Towards iris presentation attack detection with foundation models,” 2025. [Online]. Available: https://arxiv.org/abs/2501.06312

  20. [28]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec et al. , “DINOv2: Learning robust visual features without supervision,” Trans- actions on Machine Learning Research , 2024, featured Certification. [Online]. Available: https://openreview.net/forum?id=a68SUt6zFt

  21. [29]

    Iris-SAM: Iris segmentation using a foundation model,

    P. Farmanifard and A. Ross, “Iris-SAM: Iris segmentation using a foundation model,” 2024. [Online]. Available: https://arxiv.org/abs/ 2402.06497

  22. [30]

    Reproducible scaling laws for contrastive language-image learning,

    M. Cherti, R. Beaumont, R. Wightman, M. Wortsman, G. Ilharco, C. Gordon, C. Schuhmann, L. Schmidt, and J. Jitsev, “Reproducible scaling laws for contrastive language-image learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, p...

  23. [31]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” in ICML, 2021

  24. [32]

    Low-rank adaptation of pre-trained large vision models for improved lung nodule malignancy classification,

    B. P. Veasey and A. A. Amini, “Low-rank adaptation of pre-trained large vision models for improved lung nodule malignancy classification,” IEEE Open Journal of Engineering in Medicine and Biology , pp. 1–9, 2025

  25. [33]

    Yolov8: A novel object detection algorithm with enhanced performance and robustness,

    R. Varghese and S. M., “Yolov8: A novel object detection algorithm with enhanced performance and robustness,” in 2024 International Conference on Advances in Data Engineering and Intelligent Computing Systems (ADICS) , 2024, pp. 1–6

  26. [34]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” in 2017 IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) , 2017, pp. 2261–2269

  27. [35]

    EfficientNet: Rethinking model scaling for convolu- tional neural networks,

    M. Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolu- tional neural networks,” in Proceedings of the 36th Internatl. Conference on Machine Learning , ser. Proceedings of Machine Learning Research, K. Chaudhuri and R. Salakhutdinov, Eds., vol. 97. PMLR, 09–15 J...

  28. [36]

    Searching for mobilenetv3,

    A. Howard, M. Sandler, Chen et al. , “Searching for mobilenetv3,” in 2019 IEEE/CVF Internatl. Conference on Computer Vision (ICCV) , 2019, pp. 1314–1324

  29. [37]

    Deep residual learning for image recognition,

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

  30. [38]

    Swin transformer v2: Scaling up capacity and resolution,

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Dong, F. Wei, and B. Guo, “Swin transformer v2: Scaling up capacity and resolution,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 11 999–12 009. JOURNAL...

Pith tools

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