Video Transformer for Remote Identity Document Hologram Detection

Jean-Christophe Burie; Joris Voerman; Nicolas Sidere

REVIEW 2 major objections 4 minor 16 references

Video Transformer for Remote Identity Document Hologram Detection

T0 review · 2 major / 4 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A video transformer detects ID holograms from ordinary smartphone videos at near-perfect accuracy, even with limited training data.

desk verdict Solid empirical win on the only public hologram-video benchmark via a straightforward ViTransPAD adaptation; transfer to real national IDs remains the softest claim. read the letter →

arxiv 2607.11419 v1 pith:AJY6CUKX submitted 2026-07-13 cs.CV

Joris Voerman , Nicolas Sidere , Jean-Christophe Burie This is my paper · ORCID

classification cs.CV

keywords hologramdetectionremoteidentityverificationvideotransformersmartphonecaptureMIDV-Holodocumentauthenticitypresentationattack

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

Remote identity checks that rely only on a smartphone video cannot use many classic security features, because those features need special hardware. Holograms are one of the few remaining signals: they are hard to forge and produce distinctive reflections that a phone camera can capture. This paper proposes a full remote ID verification architecture and, inside it, a video transformer called ViTransHolo that decides whether a hologram is present. The model, adapted from a face presentation-attack detector, processes a short clip of cropped ID frames through an EfficientNet encoder and multi-scale self-attention layers. On the public MIDV-Holo mock-document set it raises recall by more than 26 points and accuracy by nearly 18 points over the strongest prior baseline, reaches 100 percent on a small private set of real French documents, and keeps most of that performance when the training set or the number of transformer layers is deliberately reduced. The result matters because it turns an under-used optical feature into a practical remote authenticity check that can run with modest data and compute.

What carries the argument

ViTransHolo: a short video clip is reduced to 50 equally spaced, document-cropped 224 imes224 frames, encoded by EfficientNet, then processed by a stack of multi-scale multi-head self-attention transformer layers whose frame-level scores are averaged into a single hologram-present/absent decision.

What would settle it

Train and test the same architecture on a held-out collection of genuine and forged identity documents from several different countries and capture devices never seen during training; a large drop in recall or accuracy would falsify the claimed generalisation.

Watch

Extended reading notes

Core claim

ViTransHolo, a video transformer built from an EfficientNet encoder and six multi-scale multi-head self-attention layers, detects the presence of holograms in smartphone videos of identity documents at 98.43 percent accuracy and 98.19 percent recall on MIDV-Holo, improving the best published baseline by 17.93 points in accuracy and 26.86 points in recall, and continues to work well under reduced training samples and fewer layers.

Load-bearing premise

That strong numbers obtained on two highly similar mock templates plus a 17-document French private set will transfer to the full variety of real national IDs, lighting, phones and sophisticated forgeries seen in production.

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.

Desk Editor's Note

read the letter

The paper’s real payload is the numbers: ViTransHolo (EfficientNet + six MS-MHSA layers, 50 frames) hits 98.43 % Acc / 98.19 % Recall on MIDV-Holo, beating the best published baseline by roughly +27 % Recall and +18 % Acc, with five-run averages, document-level splits, and clean per-fraud breakdowns. It also stays usable down to ~200 training videos and one transformer layer. That is a clear, reproducible advance on the only public hologram-detection set we have.

What is new is the application domain, the RIDVS system sketch, the frugality ablations (RTS/LAS), and the IML3 real-French-ID check. The architecture itself is a direct transplant of the already-published ViTransPAD face-PAD model; no new attention mechanism or loss is claimed. The authors are transparent about this.

The soft spot is exactly the one the stress-test flags, and the paper itself flags it: MIDV-Holo uses only two document classes whose hologram templates are highly similar (Fig. 4). IML3 is 17 real French IDs with a single forgery type (scan-and-reprint) and is private. Perfect scores there are therefore fragile. The frugality curves inherit the same narrow distribution, so they do not prove production robustness across national designs, lighting, devices, or sophisticated hologram forgeries. No code or public real-ID data is released. Those are real limits, but they do not erase the MIDV-Holo gains or the careful experimental hygiene.

Math and citations look clean; self-cites are to prior RIDVS modules and do not close any loop. Free parameters (frame count, depth, crop size) are standard and reported.

This is for anyone working on remote KYC, document forensics, or video transformers under data scarcity. It is not a foundational architecture paper. I would send it to peer review: the empirical result is sharp enough and the limitations are stated honestly enough that referees can decide how much weight to give the transfer claim. Worth reading if you care about the application; not essential if you only track transformer novelty.

Referee Report

2 major / 4 minor

Summary. The paper proposes a Remote Identity Document Verification System (RIDVS) for smartphone-based remote ID authentication and introduces ViTransHolo, a video transformer adapted from ViTransPAD (with EfficientNet encoder, six MS-MHSA transformer layers, and MLP head) for binary hologram presence detection. Input videos are reduced to 50 frames, document-cropped and resized to 224×224. On the public MIDV-Holo mock-document dataset the model reports 98.43% accuracy / 98.19% recall (five-run means), improving the best published baseline by +26.86% recall and +17.93% accuracy; perfect scores are obtained on the private 17-document French IML3 set. Additional controlled experiments examine per-fraud-type performance (including held-out photo-replacement), training-set size reduction (RTS), and layer ablation (LAS), claiming robustness under data and compute frugality.

Significance. Hologram detection from ordinary smartphone video is an under-explored but practically relevant security cue for remote identity verification. The empirical protocol is careful (document-level splits, five-run averages with standard deviations, explicit SotA tables, and systematic frugality ablations), and the reported gains on the only public benchmark are large. If the performance transfers beyond the two highly similar MIDV-Holo templates and the small French private set, the work would supply a usable building block for production RIDVS pipelines. The frugality results further increase practical interest. Significance is currently bounded by the narrow data distribution on which the strongest numbers rest.

major comments (2)

[V-A, Fig. 4, Tables I–V] Section V-A and Fig. 4 explicitly note that MIDV-Holo contains only two document classes whose hologram templates are highly similar. IML3 adds only 17 real French documents and a single forgery type (scan-and-reprint). Consequently the near-perfect metrics in Tables I–II and the RTS/LAS results in Tables IV–V are measured on a narrow distribution; they do not yet demonstrate that the same accuracy/recall will hold for the diversity of national hologram designs, lighting, capture devices and sophisticated forgeries that a production RIDVS must handle. The abstract and conclusion should either qualify the production claims more sharply or supply additional cross-document / cross-country evidence.
[Table I vs Table III] Table I (main SotA comparison) excludes the photo-replacement fraud class, following prior baselines, yet Table III shows a clear drop (Recall 92.78 %) when that class is tested as an unseen positive. Because photo replacement is a realistic attack that still leaves most of the hologram visible, the paper should quantify how this residual error propagates through the RIDVS decision centre and whether a reconstruction / template-matching stage (mentioned only as future work) is required before the detector can be considered production-ready.

minor comments (4)

[Abstract, §I] Abstract and Introduction contain multiple spelling/typo issues (“authentification”, “seurization”, “enlightment”, “clue in the seurization”) that should be corrected for readability.
[Fig. 2, §IV-A] Fig. 2 caption and architecture description omit the precise definition of the MS-MHSA pyramid scales and the number of attention heads; a short equation or reference to the exact ViTransPAD configuration would aid reproducibility.
[Table V] Table V header duplicates the “CTime (ms)” column label; the second column is presumably GTime.
[§IV-A] The choice of exactly 50 frames is stated without ablation; a one-sentence sensitivity note (or pointer to future work) would strengthen the experimental design claim.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; purely empirical supervised classification with held-out metrics and non-circular self-citations to prior RIDVS modules.

full rationale

The paper's central claims (near-perfect Acc/Recall on MIDV-Holo, gains over external baselines, generalization to IML3, and frugality under RTS/LAS) are obtained by standard supervised training of an adapted video transformer followed by evaluation of FPR/Recall/Acc/F1 on held-out video splits (70/15/15, document-grouped to avoid leakage, averaged over 5 runs). No quantity is defined in terms of a fitted parameter later re-presented as a prediction; no uniqueness theorem is invoked; the architecture is taken from a related-domain model (ViTransPAD) and re-validated empirically rather than assumed by construction. Self-citations ([1], [13], and overlapping authorship on [8]) describe other RIDVS components or the source architecture and do not close a logical loop around the hologram-detection numbers, which rest on independent comparison to the MIDV-Holo baseline and Triplet-loss method. The work is therefore self-contained against external benchmarks; the only residual is ordinary self-citation of prior project modules, which is non-load-bearing and warrants at most a score of 1.

Assumptions & free parameters 4 free parameters · 3 assumptions · 2 invented entities

The central empirical claim rests on standard supervised-learning assumptions plus a handful of architectural and data-handling choices that are not derived from first principles. No new physical entities are postulated; the free parameters are ordinary hyper-parameters and design decisions of the adapted transformer.

free parameters (4)

number of input frames
Fixed at 50 equally spaced frames without ablation; chosen for ‘sufficient temporal coverage while remaining tractable’.
transformer depth
Default of 6 layers; later ablated but the main claimed numbers use 6.
crop size / backbone
224×224 crops and EfficientNet encoder taken from the source ViTransPAD model; not re-tuned for holograms.
train/val/test split ratios and random seeds
70/15/15 document-level split repeated five times; exact seeds and document assignments not published.

assumptions (3)

domain assumption Holograms produce distinctive, temporally varying visual effects under ordinary smartphone illumination that are sufficient for binary presence detection.
Stated in the introduction and used as the justification for video-based detection without specialized sensors.
domain assumption Document tracking and cropping can be performed reliably enough that the subsequent classifier sees only the ID region.
Pre-processing step described in Section IV-A; failures would inject background noise that the model never saw during training.
ad hoc to paper Binary cross-entropy on frame-level labels averaged to video level is an adequate surrogate for hologram authenticity.
Training objective chosen without comparison to other temporal aggregation schemes.

invented entities (2)

ViTransHolo
purpose: Name given to the adapted video transformer for the hologram task.
Simply a re-branded and lightly modified ViTransPAD; no new architectural primitive is introduced.
RIDVS (Remote Identity Document Verification System)
purpose: Overall five-module pipeline sketch that places the hologram detector in context.
High-level system diagram; most modules are either prior work or left for future work.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Video Transformer for Remote Identity Document Hologram Detection." pith.science (2026). https://pith.science/paper/AJY6CUKX

@misc{pith2026260711419,
  author       = {Pith},
  title        = {Pith review of: Video Transformer for Remote Identity Document Hologram Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AJY6CUKX}},
  note         = {Machine review of arXiv:2607.11419}
}

read the original abstract

Remote identity authentification using Identification Documents has been a major challenge for several years. DeepFakes advent and the development of AI-guided tools helps fraudsters creating counterfeit ID Documents. Ensuring the authenticity of ID Documents has become a real clue in the seurization of remote authentification. This need is all the more pressing given the increasing digitization of administrative and transactional processes. To ensure widespread accessibility, the system should rely solely on video captured via mobile devices. In this specific context, confirming the authenticity of ID is a real challenge as many security features needs specific device like infrared sensor for instance. Among underutilized but promising security features, holographic printings hold a special place. Difficult to counterfeit, they produce distinctive visual effects according enlightment, making them both detectable in a video captured by a smartphone camera and difficult to imitate. In this paper, we propose a Remote Identity Document Verification System (RIDVS) and an approach based on a video transformer for detecting holograms in simple videos captured by smartphones. Our system is designed for a smartphone-based capture process, followed by a server-side verification. The hologram detection method builds on a robust model previously validated in a related research domain. We demonstrate that it outperforms existing SotA methods, achieving near-perfect accuracy even when trained on medium- to small-sized datasets. In particular, we report improvements of +26.86\% in Recall and +17.93\% in accuracy over the best MIDV-Holo baseline. This study includes several experiments that evaluate the model adaptation to frugality, both for training samples and computational resources.

Figures

Figures reproduced from arXiv: 2607.11419 by the authors.

**Figure 1.** Overall diagram of the proposed system for remote ID authentication. The colors correspond to specifics sub-systems : yellow for classification, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 2.** ViTransHolo model architecture. It begin with a embedding process to format input video, then an encoder (EfficientNet), followed by six [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 5.** Document classes represented in IML3. These official specimen [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: MIDV-Holo hologram templates. 2) IML3 Dataset: IML3 is a small private dataset created within the L3i laboratory, based on real French ID documents voluntarily provided by local participants. Due to the sensitive nature of the data, this dataset is not publicly availa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 1 linked inside Pith

[1]

In: IEEE/CVF Winter Conf

Al-Ghadi, M., V oerman, J., Coustaty, M., Lessard, O., Sidere, N.: Idtrust: Deep identity document quality detection with bandpass filter- ing. In: IEEE/CVF Winter Conf. on Applications of Computer Vision, W ACV 2025 - Workshops, Tucson, USA. pp. 668–675. IEEE (2025)

2025
[2]

IEEE Access10, 25114–25124 (2022)

Ay, B.: Open-set learning-based hologram verification system us- ing generative adversarial networks. IEEE Access10, 25114–25124 (2022)

2022
[3]

In: 25th Int

Chapel, M.N., Al-Ghadi, M., Burie, J.C.: Authentication of holograms with mixed patterns by direct lbp comparison. In: 25th Int. Workshop on Multimedia Signal Processing (MMSP). pp. 1–6. IEEE (2023)

2023
[4]

arXiv preprint arXiv:2010.11929 (2020)

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)

arXiv 2010
[5]

Jiao, L., Zhang, X., Liu, X., Liu, F., et al
[6]

Kada, O., Kurtz, C., van Kieu, C., Vincent, N.: Hologram detection for identity document authentication. In: Int. Conference on Pattern Recognition and Artificial Intelligence. pp. 346–357. Springer (2022)

2022
[7]

In: International Conference on Document Analysis and Recognition

Koliaskina, L., Emelianova, E., Tropin, D.V ., Popov, V ., Bulatov, K.B., Nikolaev, D.P., Arlazarov, V .V .: Midv-holo: A dataset for id document hologram detection in a video stream. In: International Conference on Document Analysis and Recognition. pp. 486–503. Springer (2023)

2023
[8]

In: IEEE International Conference on Image Processing

Ming, Z., Yu, Z., Al-Ghadi, M., Visani, M., Luqman, M.M., Burie, J.C.: Vitranspad: video transformer using convolution and self- attention for face presentation attack detection. In: IEEE International Conference on Image Processing. pp. 4248–4252. IEEE (2022)

2022

Show all 16 references

[9]

Scientific Reports12(1), 18475 (2022)

Mukundan, A., Tsao, Y .M., Lin, F.C., Wang, H.C.: Portable and low- cost hologram verification module using a snapshot-based hyperspec- tral imaging algorithm. Scientific Reports12(1), 18475 (2022)

2022
[10]

In: International Conference on Document Analysis and Recognition

Pouliquen, G., Chiron, G., Chazalon, J., G ´eraud, T., Awal, A.M.: Weakly supervised training for hologram verification in identity documents. In: International Conference on Document Analysis and Recognition. pp. 17–33. Springer (2024)

2024
[11]

Transaction on Computer Vision and Applications9(1), 9 (2017)

Soukup, D., Huber-M ¨ork, R.: Mobile hologram verification with deep learning. Transaction on Computer Vision and Applications9(1), 9 (2017)

2017
[12]

In: International conference on machine learning

Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convo- lutional neural networks. In: International conference on machine learning. pp. 6105–6114. PMLR (2019)

2019
[13]

voerman et al

V oerman, J., Al-Ghadi, M., Sidere, N., Coustaty, M., Lessard, O.: Optimizing identity documents classification in online systems: A comparative analysis: J. voerman et al. International Journal on Doc- ument Analysis and Recognition (IJDAR) pp. 1–11 (2025)

2025
[14]

In: IEEE/CVF int

Wang, D., Cui, X., Chen, X., Zou, Z., et al.: Multi-view 3d reconstruc- tion with transformers. In: IEEE/CVF int. conference on computer vision. pp. 5722–5731 (2021)

2021
[15]

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

Xie, F., Chu, L., Li, J., Lu, Y ., Ma, C.: Videotrack: Learning to track objects via video transformer. In: IEEE/CVF conference on computer vision and pattern recognition. pp. 22826–22835 (2023)

2023
[16]

Optics express29(24), 40572–40593 (2021)

Zeng, T., Zhu, Y ., Lam, E.Y .: Deep learning for digital holography: a review. Optics express29(24), 40572–40593 (2021)

2021

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.

[1] [1]

In: IEEE/CVF Winter Conf

Al-Ghadi, M., V oerman, J., Coustaty, M., Lessard, O., Sidere, N.: Idtrust: Deep identity document quality detection with bandpass filter- ing. In: IEEE/CVF Winter Conf. on Applications of Computer Vision, W ACV 2025 - Workshops, Tucson, USA. pp. 668–675. IEEE (2025)

2025

[2] [2]

IEEE Access10, 25114–25124 (2022)

Ay, B.: Open-set learning-based hologram verification system us- ing generative adversarial networks. IEEE Access10, 25114–25124 (2022)

2022

[3] [3]

In: 25th Int

Chapel, M.N., Al-Ghadi, M., Burie, J.C.: Authentication of holograms with mixed patterns by direct lbp comparison. In: 25th Int. Workshop on Multimedia Signal Processing (MMSP). pp. 1–6. IEEE (2023)

2023

[4] [4]

arXiv preprint arXiv:2010.11929 (2020)

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)

arXiv 2010

[5] [5]

Jiao, L., Zhang, X., Liu, X., Liu, F., et al

[6] [6]

Kada, O., Kurtz, C., van Kieu, C., Vincent, N.: Hologram detection for identity document authentication. In: Int. Conference on Pattern Recognition and Artificial Intelligence. pp. 346–357. Springer (2022)

2022

[7] [7]

In: International Conference on Document Analysis and Recognition

Koliaskina, L., Emelianova, E., Tropin, D.V ., Popov, V ., Bulatov, K.B., Nikolaev, D.P., Arlazarov, V .V .: Midv-holo: A dataset for id document hologram detection in a video stream. In: International Conference on Document Analysis and Recognition. pp. 486–503. Springer (2023)

2023

[8] [8]

In: IEEE International Conference on Image Processing

Ming, Z., Yu, Z., Al-Ghadi, M., Visani, M., Luqman, M.M., Burie, J.C.: Vitranspad: video transformer using convolution and self- attention for face presentation attack detection. In: IEEE International Conference on Image Processing. pp. 4248–4252. IEEE (2022)

2022