Verification of Dynamic Holographic Behavior in Identity Documents

2); (2) EPITA Research Lab (LRE); Ahmad Montaser Awal (1) ((1) IDnow Research Center; France; France); Glen Pouliquen (1; Guillaume Chiron (1); Joseph Chazalon (2); Thierry G\'eraud (2)

REVIEW 2 major objections 7 minor 21 references

Verification of Dynamic Holographic Behavior in Identity Documents

T0 review · 2 major / 7 minor · reviewed 2026-07-08 · glm-5.2

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

Pith's one-line read Hologram checker catches forged IDs without seeing fakes first

desk verdict First public dynamic hologram attack dataset plus a method that trains without seeing dynamic attacks — real contributions, with honest gaps read the letter →

arxiv 2607.06466 v1 pith:CO6BTER5 submitted 2026-07-07 cs.CV

Glen Pouliquen (1 , 2) , Joseph Chazalon (2) , Guillaume Chiron (1) , Thierry Géraud (2) , Ahmad Montaser Awal (1) ((1) IDnow Research Center , France , (2) EPITA Research Lab (LRE)

show 1 more author

France)

This is my paper

classification cs.CV

keywords hologramverificationidentitydocumentsopticallyvariabledevicesOVDpresentationattackdetectionbackgroundsubtractionpseudo-labelingfraud

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 claims that a system called HoloVerif can detect sophisticated hologram forgeries on identity documents — including types of attacks it was never trained on — by learning what genuine holographic behavior looks like rather than learning to recognize specific frauds. The method works in two stages: first, background subtraction removes the document template and bearer's photo from each video frame, isolating only the variable optical signal from the hologram; second, a classifier is trained on pseudo-labels derived from luminance changes in genuine sequences, learning the visual fingerprint of authentic holographic behavior. The authors also introduce MIDV-DynAttack, a new public dataset of 1,200 attack videos featuring handcrafted dynamic hologram forgeries — the first public dataset of its kind. On this dataset, HoloVerif achieves 61% recall on dynamic attacks and 93% AUC overall, while the prior best method (WSL) drops to 2% recall on dynamic attacks, effectively failing to detect them at all.

What carries the argument

HoloVerif uses deterministic preprocessing (per-pixel median background subtraction across a video sequence, then an HSV normalizing filter that attenuates everything except bright, highly saturated regions) to isolate the holographic signal. Training labels are generated automatically: in genuine sequences, frames with moderate luminance change ('Changing' frames) are labeled as Valid, while very bright or very dark frames are discarded. In fraudulent sequences from the existing dataset, all changing frames are labeled Non-Valid. A MobileViT-xxs classifier is trained on these pseudo-labels with asymmetric augmentation (light augmentation for Valid frames, aggressive augmentation for Non-Val

What would settle it

If a handcrafted dynamic attack (e.g., a generic holographic overlay) produces frames that pass HoloVerif's appearance classifier because the forged holographic material happens to exhibit bright, saturated, varying patterns similar to genuine OVDs, the method would fail to detect it — and the 61% recall on dynamic attacks suggests this already happens for roughly 39% of dynamic attack cases.

Watch

Extended reading notes

Core claim

The central finding is that hologram verification can be reframed as a one-class modeling problem: instead of training a detector to recognize known attack types, you train a model of genuine holographic appearance by isolating the holographic signal through background subtraction and pseudo-labeling frames where luminance changes indicate real optical variability. This model generalizes to unseen dynamic attacks because handcrafted holograms, while they may produce color variations or reflections, do not reproduce the specific appearance characteristics of genuine OVDs. The authors demonstrate this by achieving 61% recall on dynamic attacks with zero dynamic attack training samples, versus

Load-bearing premise

The pseudo-labeling step assumes that a luminance threshold (T=5) and a change-ratio threshold (1/3) correctly separate frames showing genuine holographic behavior from noise, glare, or artifacts. If these hand-tuned thresholds systematically mislabel frames — for instance, by treating glare as valid holographic behavior — the learned model would encode noise rather than genuine optical signatures. No sensitivity analysis is provided for these thresholds.

Editorial extensions

If this is right

Remote identity verification services that rely on smartphone video capture can now attempt to detect handcrafted hologram forgeries — previously an open vulnerability — without needing to collect examples of every possible forgery type.
The pseudo-labeling strategy (deriving training signal from genuine sequences alone) suggests a practical path for deploying hologram verification in production: organizations with genuine document specimens can train models without needing to source or manufacture attack samples.
The two identified failure modes — Photo Replacement and Document Swapping — define the next research frontier for hologram verification, specifically pointing toward temporal sequence modeling rather than independent frame classification.
The publicly available MIDV-DynAttack dataset establishes a standardized evaluation protocol for dynamic hologram attacks, enabling direct comparison of future methods on a problem that previously had no public benchmark.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

read the letter

The paper does two things worth your attention: it releases MIDV-DynAttack, the first public dataset of dynamic hologram attacks on identity documents (1,200 new videos), and it proposes HoloVerif, a method that detects these attacks without being trained on them. Both are genuinely new. The dataset fills a real gap — no one had published dynamic hologram attack data at this level of sophistication before. The method's core trick is clean: background subtraction isolates the holographic signal, pseudo-labels from genuine sequences provide training targets, and class-dependent augmentation teaches the classifier what valid variation looks like. The fact that code and data are public is a strong signal. The experimental protocol is solid — 5-fold cross-validation with identity grouping, 6 seeds, ablations, per-attack breakdowns in the supplementary material. The self-citation to prior work [18] is not circular; the new results are tested on genuinely unseen data. The reader's concern about pseudo-labeling thresholds (T=5, C_t > 1/3) is legitimate but probably overstated. The ablation showing 'Pseudo labels only' achieves 65% AUC on the Mix set confirms the labels carry real signal. That said, no sensitivity analysis for T or the change ratio is provided anywhere, and it should be. The per-attack results in Table F.1 are suggestive: 90% recall on Plain holo but 33% on Leaf holo hints the model learned a narrow representation. This is worth investigating but is not a load-bearing flaw — the method clearly works, just unevenly. The private rectification method is a minor reproducibility concern. The paper ships tracking information for the dataset, which partially addresses this, but the rectification pipeline itself is undisclosed despite being called 'critical.' 61% recall on dynamic attacks is honestly framed by the authors as partial progress, not a solved problem. The Document Swap failure is a structural limitation they acknowledge openly. This paper is for document forensics and KYC researchers. It deserves a serious referee who can push on the threshold sensitivity question and the rectification disclosure. The core contributions hold up.

Referee Report

2 major / 7 minor

Summary. This paper introduces MIDV-DynAttack, a new public test-only dataset of 1,200 video sequences extending MIDV-Holo with realistic static and dynamic hologram attacks on identity documents. The authors also propose HoloVerif, a verification method that uses background subtraction to isolate holographic signals, pseudo-labels frames based on luminance-change heuristics, and trains a frame-level classifier with class-dependent augmentations. The method is trained without dynamic attack samples and evaluated on unseen attacks from MIDV-DynAttack. Results show HoloVerif achieves 61% recall on dynamic attacks and 93% AUC on the combined test set, outperforming prior methods including the authors' own WSL approach, which drops to 2% recall on dynamic attacks. The experimental protocol is thorough: 5-fold cross-validation with identity grouping, 6 random seeds, ablation studies, and per-attack breakdowns. Code and dataset are publicly available.

Significance. The paper makes a solid contribution to the document forensics community. The MIDV-DynAttack dataset fills a genuine gap: no prior public dataset featured dynamic hologram attacks of this sophistication, and the authors correctly identify that existing methods have never been evaluated against such attacks. The finding that WSL (the prior state of the art) collapses to 2% recall on dynamic attacks is itself a valuable community alert. The HoloVerif method is well-motivated: by subtracting the per-sequence background median, it removes document-specific content (including the bearer's photo) and forces the classifier to attend to holographic residuals. The pseudo-labeling heuristic is a creative way to avoid per-frame annotation. The ablation study (Table 2) is informative, particularly the 'Pseudo labels only' and 'No augmentations' rows, which isolate the contributions of each component. The public release of code and dataset is a significant strength that enables reproducibility and future benchmarking. The per-attack breakdown (Table F.1) provides useful diagnostic granularity. The paper is honest about limitations, including the Document Swap vulnerability and the lack of new-g

major comments (2)

Section 4.2, pseudo-labeling thresholds: The luminance threshold T=5 and change-ratio threshold C_t > 1/3 are the sole mechanism generating positive training labels for the frame classifier, yet no sensitivity analysis is provided anywhere in the paper or supplementary. If these thresholds systematically include non-holographic residuals (e.g., glare after background subtraction, motion artifacts from imperfect alignment) or exclude genuine but subtle holographic variation, the learned model would encode spurious patterns. Table F.1 provides circumstantial evidence that this may be occurring: recall on 'Plain holo' is 90% but drops to 33% on 'Leaf holo,' suggesting the model has learned a narrow representation. A simple sweep over T in {3, 5, 7, 10} and the ratio threshold in {0.2, 0.33, 0.5}, reporting the resulting AUC on the Mix set, would address this concern and is feasible given 5×
Table 2, ablation rows: The jump from 'Pseudo labels only' (65% AUC on Mix) to full HoloVerif (93% AUC) is 28 points, while the 'No augmentations' ablation achieves 87% AUC. This means augmentations contribute roughly 22 points on top of the learned classifier, but the pseudo-label quality contribution is only ~2 points (87 minus 65 is not the right decomposition because 'No augmentations' still uses the trained classifier). The paper does not clearly disentangle pseudo-label quality from augmentation effects. A cleaner ablation would vary T and the ratio threshold while holding augmentation fixed, or report pseudo-label purity (e.g., precision of the 'Valid' label against a small manually annotated subset). Without this, it is unclear whether the pseudo-labeling mechanism is robust or whether the augmentations are compensating for label noise.

minor comments (7)

Section 4.1: The paper states the preprocessing step 'does not require any specific parameters,' but the normalizing filter N involves a max over all timestamps and pixels (Eq. 1), and the background subtraction uses a per-pixel median. While these are parameter-free in the traditional sense, the claim could be misread as implying no design choices. Consider rephrasing to 'does not require hand-tuned parameters.'
Section 5.1: The private rectification method is described at a high level (local-descriptor matching, temporal coherence, trajectory completion) but key details (which local descriptors, matching thresholds, how 'unreliable' detections are identified) are omitted. The paper itself flags rectification quality as 'critical.' Since tracking information is provided in the dataset, this is a secondary reproducibility concern, but a brief note on the descriptor type and any thresholds would help readers assess the alignment quality.
Table 2: The 'MIDV-Holo' method row reports '3‡ ±3' for Dynamic recall with a footnote 'low Dynamic scores by design.' This is clear, but the dagger symbol '†' in the ablation rows is defined as 'showing strong negative impact' which is somewhat vague. Consider replacing with a quantitative note (e.g., 'p < 0.05 vs. full HoloVerif') or making it more prominent.
Table F.1: The 'Laser light' and 'LED light' rows show 0±0 recall for MIDV-Holo and WSL, but 100±0 for HoloVerif. The zero variance is surprising for 50-video categories. A brief note explaining whether this reflects perfect detection or a small-sample artifact (and whether these attacks are trivially distinguishable as non-holographic).
Section 3: The dataset description mentions 1,200 videos but the breakdown (750 static + 450 dynamic) does not explicitly state how many unique document templates and identities are covered. Table 1 mentions 100 IDs and 20 doc types, but a sentence clarifying the template/identity structure of MIDV-DynAttack (e.g., 'N templates × 5 identities') would help.
Figure 3: The bottom-row zoomed panels are small and the axis labels overlap. Consider enlarging or providing separate figures.
Supplementary Section D.2: The AUC computation uses 100 evenly distributed thresholds from the MIDV-Holo Vanilla ROC, then applies them to MIDV-DynAttack. This is a reasonable design, but a one-sentence justification for why this is preferable to computing AUC directly on the combined set would clarify the cross-dataset AUC.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. Both major points concern the pseudo-labeling mechanism—its sensitivity to threshold choices and its interaction with augmentations in the ablation. We agree both are valid concerns and will address them in the revision.

read point-by-point responses

Referee: Section 4.2, pseudo-labeling thresholds: The luminance threshold T=5 and change-ratio threshold C_t > 1/3 are the sole mechanism generating positive training labels for the frame classifier, yet no sensitivity analysis is provided anywhere in the paper or supplementary. If these thresholds systematically include non-holographic residuals (e.g., glare after background subtraction, motion artifacts from imperfect alignment) or exclude genuine but subtle holographic variation, the learned model would encode spurious patterns. Table F.1 provides circumstantial evidence that this may be occurring: recall on 'Plain holo' is 90% but drops to 33% on 'Leaf holo,' suggesting the model has learned a narrow representation. A simple sweep over T in {3, 5, 7, 10} and the ratio threshold in {0.2, 0.33, 0.5}, reporting the resulting AUC on the Mix set, would address this concern and is feasible given 5×

Authors: We agree that a sensitivity analysis over the pseudo-labeling thresholds is missing and should be added. The referee is correct that T and C_t are the sole mechanism generating positive training labels, so their robustness matters for the credibility of the method. We will run the proposed sweep over T ∈ {3, 5, 7, 10} and ratio threshold ∈ {0.2, 0.33, 0.5}, reporting AUC on the Mix set (and per-attack recall where informative), and add the results to the supplementary material. We expect this to also shed light on the per-attack variance the referee notes. We would add one nuance: the 90% vs. 33% gap between 'Plain holo' and 'Leaf holo' may not be solely attributable to threshold sensitivity. The Leaf holo material produces distinctive polygonal light patterns that are visually closer to genuine OVD diffraction patterns than the homogeneous Plain holo material, making it inherently harder to reject regardless of labeling thresholds. The sensitivity sweep will help disentangle these two factors. revision: yes
Referee: Table 2, ablation rows: The jump from 'Pseudo labels only' (65% AUC on Mix) to full HoloVerif (93% AUC) is 28 points, while the 'No augmentations' ablation achieves 87% AUC. This means augmentations contribute roughly 22 points on top of the learned classifier, but the pseudo-label quality contribution is only ~2 points (87 minus 65 is not the right decomposition because 'No augmentations' still uses the trained classifier). The paper does not clearly disentangle pseudo-label quality from augmentation effects. A cleaner ablation would vary T and the ratio threshold while holding augmentation fixed, or report pseudo-label purity (e.g., precision of the 'Valid' label against a small manually annotated subset). Without this, it is unclear whether the pseudo-labeling mechanism is robust or whether the augmentations are compensating for label noise.

Authors: The referee correctly identifies that our ablation does not cleanly disentangle pseudo-label quality from augmentation effects. The 'Pseudo labels only' row uses the hardcoded decision function (no trained classifier), while 'No augmentations' uses the trained classifier without augmentation, so the 87−65=22 point gap conflates classifier learning with pseudo-label quality rather than isolating either. We accept this criticism. We will address it in two ways: (1) We will report pseudo-label purity—precision and recall of the 'Valid'/'Non-Valid' labels against a manually annotated subset of frames (approximately 500 frames sampled across Legit and Non-legit sequences, covering Bright, Changing, and discarded frames). This directly measures label noise without confounding by augmentation. (2) We will add the threshold sweep from the first comment as an additional ablation axis, holding augmentation fixed, which will show whether pseudo-label quality (as controlled by T and C_t) affects final performance independently of augmentation. Together these should clarify whether augmentations are compensating for label noise or genuinely expanding the model's effective training distribution. We will revise the ablation discussion in Section 5.3 accordingly. revision: yes

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is independently testable on the new MIDV-DynAttack test set, and self-citations are used for protocol/dataset continuity, not as load-bearing logical premises.

full rationale

The paper's central claim — that HoloVerif, trained without dynamic attack samples, achieves improved detection of dynamic hologram attacks — is evaluated against MIDV-DynAttack, a dataset explicitly designed as test-only and not used during training or calibration (Section 3: 'all MIDV-DynAttack is intended to be used as a test set, with every attack not seen during training'). The pseudo-labeling procedure (Section 4.2) generates training labels from genuine sequences using fixed thresholds (T=5, change ratio 1/3), but these labels are derived from the MIDV-Holo training split, not from the test data, so the evaluation is not circularly forced. The self-citation to [18] (Pouliquen et al.) provides the evaluation protocol, the WSL baseline, and the Direct Classifier baseline; these are used for comparison and methodological continuity, not as premises that define the present result. The WSL method [18] is treated as a baseline to be outperformed, not as a load-bearing assumption. The pseudo-labeling thresholds lack sensitivity analysis (a correctness concern), and the 'Pseudo labels only' ablation (65% AUC on Mix) confirms the pseudo-labels carry genuine signal rather than being tautological with the test set. No step in the derivation chain reduces to its own inputs by construction. The concerns about threshold robustness and per-attack recall variability are correctness risks, not circularity.

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

The paper introduces no new physical entities, particles, forces, or mathematical objects. The free parameters are all standard hyperparameters (thresholds, augmentation magnitudes) typical of machine learning pipelines. The axioms are domain-specific physical assumptions about hologram optics and document capture, not ad hoc inventions.

free parameters (4)

Luminance threshold T = 5
Used in pseudo-labeling (Section 4.2) to threshold preprocessed image luminance. Appears hand-set with no justification or sensitivity analysis.
Change ratio threshold = 1/3
Frames with C_t > 1/3 are labeled 'Bright', C_t < 1e-4 are 'Dark'. Hand-set with no sensitivity analysis.
Decision threshold = calibrated on validation set
Threshold on ratio of Valid frames for sequence verdict, calibrated to maximize F-score on validation set. Standard practice but a fitted parameter.
Augmentation parameters = various (brightness ±0.3, contrast ±0.3, etc.)
Class-dependent augmentation magnitudes in Section C of supplementary. Hand-tuned.

assumptions (3)

domain assumption Genuine holographic behavior produces bright, highly saturated color variations when the document is tilted, distinguishable from static template artifacts.
Section 4.1: the HSV normalizing filter is designed on the assumption that 'only regions exhibiting high brightness and high color saturation — precisely the optical signature of genuine holograms — remain prominent.' This is the physical basis for the entire preprocessing pipeline.
domain assumption Frame-level holographic behavior can be learned independently of temporal ordering, i.e., the set of valid frames is sufficient without modeling transitions.
Section 4.3: the final decision is a threshold on the ratio of Valid frames, treating frames as an unordered set. The paper itself acknowledges this is a limitation in Section 5.3: 'improve hologram modeling with a temporal aspect... instead of processing it as a set of frames.'
domain assumption The rectification process produces sufficiently accurate frame alignment for background subtraction to work.
Section 4.1: 'The quality of the rectification process (not described in this paper), which precisely aligns every frame of the sequence, is critical here.' The method's correctness depends on this unstated external component.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Verification of Dynamic Holographic Behavior in Identity Documents." pith.science (2026). https://pith.science/paper/CO6BTER5

@misc{pith2026260706466,
  author       = {Pith},
  title        = {Pith review of: Verification of Dynamic Holographic Behavior in Identity Documents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CO6BTER5}},
  note         = {Machine review of arXiv:2607.06466}
}

read the original abstract

This paper addresses the remote verification of the authenticity of Optically Variable Devices (commonly known as holograms) on identity documents. Typically placed over the cardholder's photo, these devices provide strong and easily verifiable security for human inspection but pose challenges for automated verification. Existing approaches easily cover static frauds (e.g. paper photocopy) and can be evaluated for such, but their capacity to detect real, dynamic fraud cases (e.g. handcrafted hologram) has not been evaluated to date because of the lack of public datasets. Furthermore, they are usually trained to detect known attack types, and few of them can generalize to new, unseen attacks. This work features three contributions to address these limitations: 1) a new public dataset, MIDV-DynAttack, which extends the existing MIDV-Holo dataset with realistic, static and dynamic attacks against identity document specimens, tripling the number of attack samples compared to the original dataset, 2) a novel verification method which can assess the authenticity of a specific hologram thanks to the analysis of its dynamic behavior and appearance, can be trained without dynamic attack samples, and exhibits new state-of-the-art performance, 3) a benchmark of existing approaches which follows a clear evaluation protocol and emphasizes the inability of other approaches to deal with dynamic attacks, as well as new challenging attacks to deal with. Code and dataset are publicly available at https://github.com/EPITAResearchLab/pouliquen.25.icdar.

Figures

Figures reproduced from arXiv: 2607.06466 by the authors.

**Figure 1.** Our proposed dataset MIDV-DynAttack extends the original MIDV-Holo dataset with 1200 new attack videos. MIDV-DynAttack is designed for testing purposes over unseen attacks only, and not for model training or calibration. limited features detectable by commodity cameras on smartphones, leading to potential creative attacks that can bypass existing systems. To address these attacks, two main directions have been expl… view at source ↗

**Figure 2.** Overview of the proposed method. Its preprocessing step is common to both training and inference phases. The frame classifier is trained over pseudo-labels generated from a generic holographic content detector. The final decision is derived from individual frame classifications. 4 Proposed Method for OVD Verification The proposed method processes a video sequence of an OVD (i.e., rectified and pre-cropped to the Re… view at source ↗

**Figure 3.** ROC curve of selected methods for MIDV-Holo “Vanilla” (left) and MIDVDynAttack (right) test sets, after calibration on a separate set. Bottom row shows zoomed regions for either low false positive rate (easier to use), or high true positive rate (safer) regimes. A True Positive is a fraud sample properly detected. The transparent area around each curve indicates the variance (±σ) observed over 5 folds (× 6 seeds fo… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[18]

Weakly Supervised Training for Hologram Verification in Identity Documents

Glen Pouliquen, Guillaume Chiron, Joseph Chazalon, Thierry Géraud, and Ahmad Montaser Awal. “Weakly Supervised Training for Hologram Verification in Identity Documents”. In:Proceedings of the 18th Interna- tional Conference on Document Analysis and Recognition (ICDAR). 2024, pp. 17–33.doi:10.1007/978-3-031-70533-5_2

work page doi:10.1007/978-3-031-70533-5_2 2024
[1]

Smart- Doc 2017 Video Capture: Mobile Document Acquisition in Video Mode

J. Chazalon, P. Gomez-Krämer, J.-C. Burie, M. Coustaty, S. Eskenazi, M. Luqman, N. Nayef, M. Rusiñol, N. Sidère, and J.-M. Ogier. “Smart- Doc 2017 Video Capture: Mobile Document Acquisition in Video Mode”. In:Proceedings of the 14th IAPR International Conference on Document Analysis and Recognition (ICDAR). Vol. 04. 2017, pp. 11–16.doi:10 . 1109/ICDAR.2017.306

work page 2017
[2]

Mobile hologram verification with deep learning

Daniel Soukup and Reinhold Huber-Mörk. “Mobile hologram verification with deep learning”. In:IPSJ Transactions on Computer Vision and Ap- plications9 (2017).doi:10.1186/s41074-017-0022-7

work page doi:10.1186/s41074-017-0022-7 2017
[3]

MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream

VladimirV.Arlazarov,KonstantinBulatov,TimofeyChernov,andVladimir L. Arlazarov. “MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream”. In:Computer Optics43.5 (2019).doi:10.18287/2412-6179-2019-43-5-818-824

work page doi:10.18287/2412-6179-2019-43-5-818-824 2019
[4]

MIDV- 2019: Challenges of the modern mobile-based document OCR

Konstantin Bulatov, Daniil Matalov, and Vladimir V. Arlazarov. “MIDV- 2019: Challenges of the modern mobile-based document OCR”. In:Pro- ceedings of the twelfth International Conference on Machine Vision (ICMV 2019). 2020, p. 64.doi:10.1117/12.2558438

work page doi:10.1117/12.2558438 2019
[5]

BIDDataset: a challenge dataset for document processing tasks

ÁlyssonSoares,RicardodasNevesJunior,andByronBezerra.“BIDDataset: a challenge dataset for document processing tasks”. In:Proceedings of the XXXIII Conference on Graphics, Patterns and Images. 2020, pp. 143–146. doi:10.5753/sibgrapi.est.2020.12997

work page doi:10.5753/sibgrapi.est.2020.12997 2020
[6]

MIDV-LAIT: A Challenging Dataset for Recognition of IDs with Perso-Arabic, Thai, and Indian Scripts

YuliaChernyshova,EkaterinaEmelianova,AlexanderSheshkus,andVladimir Arlazarov. “MIDV-LAIT: A Challenging Dataset for Recognition of IDs with Perso-Arabic, Thai, and Indian Scripts”. In:Proceedings of the 16th International Conference Document Analysis and Recognition.2021,pp.258– 272.doi:10.1007/978-3-030-86331-9_17

work page doi:10.1007/978-3-030-86331-9_17 2021
[7]

Open-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks

Betul Ay. “Open-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks”. In:IEEE Access10 (2022), pp. 25114– 25124.doi:10.1109/ACCESS.2022.3155870

work page doi:10.1109/access.2022.3155870 2022

Show all 21 references

[8]

MIDV-2020:AComprehensiveBenchmarkDatasetforIden- tityDocumentAnalysis

Konstantin Bulatov, Ekaterina Emelianova, Daniil Tropin, Natalya Sko- ryukina, Yulia Chernyshova, Alexander Sheshkus, Sergey Usilin, Zuheng Ming,Jean-ChristopheBurie,MuhammadMuzzamilLuqman,andVladimir V.Arlazarov.“MIDV-2020:AComprehensiveBenchmarkDatasetforIden- tityDocumentAn...

2020
[9]

Hologram Detection for Identity Document Authentication

Oumayma Kada, Camille Kurtz, Cuong van Kieu, and Nicole Vincent. “Hologram Detection for Identity Document Authentication”. In:Proceed- ings of the 3rd International Conference on Pattern Recognition and Ar- tificial Intelligence. 2022, pp. 346–357.doi:10.1007/978-3-031-09037- 0_29

2022 doi
[10]

Document Liveness Challenge Dataset Verification of Dynamic Holographic Behavior in Identity Documents 17 (DLC-2021)

Dmitry V. Polevoy, Irina V. Sigareva, Daria M. Ershova, Vladimir V. Ar- lazarov, Dmitry P. Nikolaev, Zuheng Ming, Muhammad Muzzamil Luq- man, and Jean-Christophe Burie. “Document Liveness Challenge Dataset Verification of Dynamic Holographic Behavior in Identity Documents 17 (...

2021
[11]

Au- thentication of Holograms with Mixed Patterns by Direct LBP Compar- ison

Marie-Neige Chapel, Musab Al-Ghadi, and Jean-Christophe Burie. “Au- thentication of Holograms with Mixed Patterns by Direct LBP Compar- ison”. In:Proceedings of the 25th International Workshop on Multime- dia Signal Processing. 2023, pp. 1–6.doi:10 . 1109 / MMSP59012 . 2023 . 10337669

2023
[12]

Guilloche Detection for ID Authentication: A Dataset and Baselines

Musab Al-Ghadi, Zuheng Ming, Petra Gomez-Krämer, Jean-Christophe Burie, Mickaël Coustaty, and Nicolas Sidere. “Guilloche Detection for ID Authentication: A Dataset and Baselines”. In:Proceedings of the 25th In- ternational Workshop on Multimedia Signal Processing. 2023, pp. 1–...

2023 doi
[13]

MIDV-Holo: A Dataset for ID Document Hologram Detection in a Video Stream

L. I. Koliaskina, E. V. Emelianova, D. V. Tropin, V. V. Popov, K. B. Bulatov, D. P. Nikolaev, and V. V. Arlazarov. “MIDV-Holo: A Dataset for ID Document Hologram Detection in a Video Stream”. In:Proceedings of the 17th International Conference on Document Analysis and Recognit...

2023 doi
[14]

KID34K: A Dataset for Online Identity Card Fraud Detection

Eun-JuPark,Seung-YeonBack,JeonghoKim,andSimonS.Woo.“KID34K: A Dataset for Online Identity Card Fraud Detection”. In:Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2023, pp. 5381–5385.doi:10.1145/3583780.3615122

2023 doi
[15]

Syn- thetic dataset of id and travel documents

Carlos Boned, Maxime Talarmain, Nabil Ghanmi, Guillaume Chiron, San- ket Biswas, Ahmad Montaser Awal, and Oriol Ramos Terrades. “Syn- thetic dataset of id and travel documents”. In:Scientific Data11.1 (2024), p. 1356.doi:10.1038/s41597-024-04160-9

2024 doi
[16]

Julien Lerouge, Guillaume Betmont, Thomas Bres, Evgeny Stepankevich, and Alexis Bergès.DocXPand-25k: a large and diverse benchmark dataset for identity documents analysis. 2024. eprint:arXiv/2408.01690

2024 arXiv
[17]

DocLightDetect: A New Algorithm for Occlusion Classifica- tion in Identification Documents

Ricardo Batista das Neves Junior, Byron Leite Dantas Bezerra, and Cleber Zanchettin. “DocLightDetect: A New Algorithm for Occlusion Classifica- tion in Identification Documents”. In:Proceedings of the 16th IAPR Inter- national Workshop on Document Analysis Systems (DAS). 2024,...

2024 doi
[19]

A Novel and Responsible Dataset for Face Presentation Attack Detection on Mobile Devices

Nathan Ramoly, Alain Komaty, Vedrana Krivokuća Hahn, Lara Younes, Ahmad-Montaser Awal, and Sébastien Marcel. “A Novel and Responsible Dataset for Face Presentation Attack Detection on Mobile Devices”. In: Proceedings of the IEEE International Joint Conference on Biometrics (IJ...

2024 doi
[20]

First Competition on Pre- 18 Pouliquen et al. sentation Attack Detection on ID Card

Juan E. Tapia, Naser Damer, Christoph Busch, Juan M. Espin, Javier Barrachina, Alvaro S. Rocamora, and et al. “First Competition on Pre- 18 Pouliquen et al. sentation Attack Detection on ID Card”. In:Proceedings of the IEEE In- ternational Joint Conference on Biometrics (IJCB)...

2024 doi
[21]

IDNet: A Novel Dataset for Identity Document Analysis and Fraud Detection

Lulu Xie, Yancheng Wang, Hong Guan, Soham Nag, Rajeev Goel, Niran- jan Swamy, Yingzhen Yang, Chaowei Xiao, Jonathan Prisby, Ross Ma- ciejewski, and Jia Zou. “IDNet: A Novel Dataset for Identity Document Analysis and Fraud Detection”. In:Proceedings of the IEEE International Co...

2024

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.

[1] [18]

Weakly Supervised Training for Hologram Verification in Identity Documents

Glen Pouliquen, Guillaume Chiron, Joseph Chazalon, Thierry Géraud, and Ahmad Montaser Awal. “Weakly Supervised Training for Hologram Verification in Identity Documents”. In:Proceedings of the 18th Interna- tional Conference on Document Analysis and Recognition (ICDAR). 2024, pp. 17–33.doi:10.1007/978-3-031-70533-5_2

work page doi:10.1007/978-3-031-70533-5_2 2024

[2] [1]

Smart- Doc 2017 Video Capture: Mobile Document Acquisition in Video Mode

J. Chazalon, P. Gomez-Krämer, J.-C. Burie, M. Coustaty, S. Eskenazi, M. Luqman, N. Nayef, M. Rusiñol, N. Sidère, and J.-M. Ogier. “Smart- Doc 2017 Video Capture: Mobile Document Acquisition in Video Mode”. In:Proceedings of the 14th IAPR International Conference on Document Analysis and Recognition (ICDAR). Vol. 04. 2017, pp. 11–16.doi:10 . 1109/ICDAR.2017.306

work page 2017

[3] [2]

Mobile hologram verification with deep learning

Daniel Soukup and Reinhold Huber-Mörk. “Mobile hologram verification with deep learning”. In:IPSJ Transactions on Computer Vision and Ap- plications9 (2017).doi:10.1186/s41074-017-0022-7

work page doi:10.1186/s41074-017-0022-7 2017

[4] [3]

MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream

VladimirV.Arlazarov,KonstantinBulatov,TimofeyChernov,andVladimir L. Arlazarov. “MIDV-500: A Dataset for Identity Documents Analysis and Recognition on Mobile Devices in Video Stream”. In:Computer Optics43.5 (2019).doi:10.18287/2412-6179-2019-43-5-818-824

work page doi:10.18287/2412-6179-2019-43-5-818-824 2019

[5] [4]

MIDV- 2019: Challenges of the modern mobile-based document OCR

Konstantin Bulatov, Daniil Matalov, and Vladimir V. Arlazarov. “MIDV- 2019: Challenges of the modern mobile-based document OCR”. In:Pro- ceedings of the twelfth International Conference on Machine Vision (ICMV 2019). 2020, p. 64.doi:10.1117/12.2558438

work page doi:10.1117/12.2558438 2019

[6] [5]

BIDDataset: a challenge dataset for document processing tasks

ÁlyssonSoares,RicardodasNevesJunior,andByronBezerra.“BIDDataset: a challenge dataset for document processing tasks”. In:Proceedings of the XXXIII Conference on Graphics, Patterns and Images. 2020, pp. 143–146. doi:10.5753/sibgrapi.est.2020.12997

work page doi:10.5753/sibgrapi.est.2020.12997 2020

[7] [6]

MIDV-LAIT: A Challenging Dataset for Recognition of IDs with Perso-Arabic, Thai, and Indian Scripts

YuliaChernyshova,EkaterinaEmelianova,AlexanderSheshkus,andVladimir Arlazarov. “MIDV-LAIT: A Challenging Dataset for Recognition of IDs with Perso-Arabic, Thai, and Indian Scripts”. In:Proceedings of the 16th International Conference Document Analysis and Recognition.2021,pp.258– 272.doi:10.1007/978-3-030-86331-9_17

work page doi:10.1007/978-3-030-86331-9_17 2021

[8] [7]

Open-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks

Betul Ay. “Open-Set Learning-Based Hologram Verification System Using Generative Adversarial Networks”. In:IEEE Access10 (2022), pp. 25114– 25124.doi:10.1109/ACCESS.2022.3155870

work page doi:10.1109/access.2022.3155870 2022