Exploring Light-Weight Object Recognition for Real-Time Document Detection

David Menotti; Lucas Wojcik; Luiz Coelho; Roger Granada

REVIEW 2 major objections 4 minor 22 references

Exploring Light-Weight Object Recognition for Real-Time Document Detection

T0 review · 2 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A lightweight license-plate detector, retrained on synthetic ID images, can run in milliseconds and match state-of-the-art document rectification in end-to-end OCR quality.

desk verdict Adapting a tiny license-plate detector to documents gives a real speed win, but the 'just as good as SOTA' OCR claim rests on a single cleaned 236-image synthetic run. read the letter →

arxiv 2509.06246 v1 pith:YLSGQBDZ submitted 2025-09-07 cs.CV

Lucas Wojcik , Luiz Coelho , Roger Granada , David Menotti This is my paper · ORCID

classification cs.CV

keywords documentdetectionrectificationreal-timeOCRIWPOD-NetlightweightneuralnetworkdataaugmentationsyntheticIDdocumentsLevenshteinmetric

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that fast document detection for OCR does not require a large, document-specific model. It adapts IWPOD-Net, a compact fully convolutional network built for license plates, retrains it from scratch on a synthetic Brazilian ID dataset, and tunes its data augmentation. On the paper's end-to-end OCR measure, the 1.8M-parameter model matches or beats larger rotated-object detectors and a dedicated skew estimator while running in a few milliseconds. The practical message is that rectification does not need to be geometrically perfect: once text is upright and visible, a strong OCR engine does the rest.

What carries the argument

IWPOD-Net's affine-warp output: the network encodes each detected object as six affine-transformation parameters that map a canonical square onto the document polygon in the image, with two independent shallow subnetworks predicting object probability and localization so conflicting gradients do not mix. The adapted pipeline also uses the model's online augmentation—random crop, randomized roll/pitch/yaw, and photometric changes—retuned so a few hundred synthetic examples do not overfit. These pieces make a single-class, multi-object detector designed for license plates transferable to ID documents with the same canonical ratio.

What would settle it

Repeat the Gemini OCR evaluation, say ten times with the same prompt, on the rectified images from each method and compare the spread to the 0.5-point margins separating IWPOD (97.73), YOLO11 (97.17), and GT boxes (97.43). Then run the same evaluation on the unfiltered NBID folds or on real ID-card photos. If the run-to-run spread exceeds those margins, or if IWPOD's score drops below the other rectification methods on real photos, the equality claim is refuted.

Watch

Extended reading notes

Core claim

On a 236-image subset of NBID, the best-tuned IWPOD-Net reaches 95.53 mean IoU with 1.8M parameters, at 12.56 ms per image on CPU and 4.96 ms on GPU. Under a Levenshtein-based OCR score with Gemini 2.0, it scores 97.73, tied with the Jdeskew skew estimator, slightly above ground-truth rectangles (97.43), YOLO11-tiny (97.17), and RTMDet-tiny (97.16); sending the raw photo scores 89.04. The same pattern holds with EasyOCR, where all rectification methods land near 89. The paper argues this means a license-plate detector can be repurposed for documents because they share a fixed canonical aspect ratio, that tuned 3D-rotation augmentation is what prevents overfitting on a few hundred synthetic i

Load-bearing premise

The claim depends on a single Gemini run over 236 synthetic images that were filtered to remove severe distortion; if OCR scores vary from run to run, or if synthetic images do not represent real photos, a half-point margin between methods is not conclusive.

Editorial extensions

If this is right

A mobile OCR pipeline can replace heavier rectification modules with a detector that runs in about 5 ms on GPU and 12 ms on CPU while keeping retrieval quality.
Detector rankings by IoU do not necessarily match rankings by OCR quality, so end-to-end OCR plus latency is the more relevant metric for document pipelines.
Raw photos sent directly to a strong multimodal OCR model still lose about 8-9 points versus any rectification method, so preprocessing remains worthwhile even as OCR models improve.
Small fully convolutional networks can be trained from scratch on a few hundred synthetic document images, provided the augmentation gives enough geometric variety.

Reading between the lines

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

(Editorial inference) The same canonical-square encoding should transfer to other fixed-shape documents, but variable aspect ratios such as invoices would require multiple anchors or a different parameterization.
(Editorial inference) Since Jdeskew and IWPOD tie exactly on the Gemini score, the experiment may be hitting the OCR model's ceiling; a more discriminating test would use harder documents or several OCR engines.
(Editorial inference) A direct test of augmentation components on real photos could show whether roll, pitch, or yaw drives the gain; if roll dominates, deployers of rotation-stable documents could simplify the augmentation and cut training time.

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

Short version: the efficiency claim is real; the OCR parity claim is plausible but not yet nailed down. The paper deserves a serious referee.

What's actually new: they take IWPOD-Net, a license-plate detector, retrain it from scratch on NBID, run a careful augmentation ablation, a cross-dataset check with MIDV, and evaluate end-to-end OCR using a normalized Levenshtein score. The speed numbers in Table III are the real result: 1.8M parameters, 12.56 ms CPU and 4.96 ms GPU, versus YOLO11-tiny at 2.6M/50.74/6.38 and RTMDet-tiny at 4.9M/1.24s/18.04. Detection IoU is also better (95.53 vs 89.61 and 93.7). That is a genuine efficiency win for a mobile-first pipeline, and the code and folds are public.

Soft spots, in order of size. The 'just as good as SOTA' OCR claim rests on a single Gemini-2.0 run over 236 NBID images with no variance reported. On top of that, the authors manually removed 'a few instances that featured severe document distortion' (Section III-A) — precisely the hard cases a lightweight detector might fail on. The OCR differences are a few tenths of a point (IWPOD 97.73, GT box 97.43), which could easily be run-to-run noise. The paper honestly flags this: the conclusion says real-data evaluation is future work and that synthetic data has limitations for OCR fidelity. So I read the parity claim as suggestive, not established. The 'novel OCR metric' is a standard normalized edit distance; fine as an evaluation tool, no new mechanism. The cross-dataset numbers are low (60–76 IoU), but the paper reports them without spin, so I don't hold that against it.

I disagree with any reading that calls this circular or a fitting exercise. The model is trained on detection, the OCR metric is independent, and the Levenshtein score is an explicit measure. The main weakness is statistical: one OCR run on a filtered subset.

Who this is for: people working on document detection/rectification pipelines for mobile OCR, and anyone interested in how much rectification quality OCR actually needs. It's a solid applied benchmark. Send it to peer review — a referee should push for repeated OCR runs, a less filtered test set, and ideally one real-data sanity check.

Referee Report

2 major / 4 minor

Summary. The paper adapts IWPOD-Net, a license-plate detector, to detect ID documents in the synthetic NBID dataset. It conducts an augmentation ablation, cross-dataset experiments with MIDV, and compares rectification methods by feeding rectified documents to Gemini-2.0 and EasyOCR, evaluating with a normalized Levenshtein-based OCR score. The main claim is that the tuned IWPOD-Net is faster and has OCR quality comparable to state-of-the-art detectors while being much smaller (1.8M parameters, 12.56 ms CPU, 4.96 ms GPU).

Significance. If the parity claim holds, the paper demonstrates a practically valuable result: a very small detector can replace larger rectification models without degrading OCR fidelity, enabling real-time document processing on mobile-class hardware. The paper's strengths include public code and fold splits, a 10-fold cross-validation protocol, explicit augmentation ablations, a clearly defined OCR metric, and a head-to-head comparison against multiple detectors and a skew estimator. The primary weakness is that the central 'just as good' claim rests on a single, filtered evaluation run with very small OCR score differences, leaving the parity conclusion statistically unverified.

major comments (2)

[Section V, Table III] The claim that IWPOD-Net is 'just as good' as state-of-the-art approaches in OCR quality is not adequately supported. The Gemini scores are based on a single run, and the differences between IWPOD (97.73), YOLO11 (97.17), and RTMDet (97.16) are only 0.5–0.6 points. More tellingly, ground-truth box rectification scores 97.43, lower than IWPOD's 97.73, which is inconsistent with the assumption that better rectification should not hurt OCR. This pattern strongly suggests run-to-run stochasticity in the OCR engine or in the entity-matching prompt. The authors should report means and variances over multiple Gemini runs, and perform a paired significance test (e.g., Wilcoxon signed-rank) between IWPOD and each competitor before making a parity claim.
[Section III-A, Section VI] The evaluation that underpins the main claim uses a filtered synthetic subset: the authors manually removed instances with severe document distortion (Section III-A), leaving only 236 images, and the paper explicitly defers real-data evaluation to future work ('we plan on performing a similar study involving real data'). The removed instances are precisely the hard cases where a lightweight detector might fail, and Table II shows that the model's cross-dataset generalization is poor (NBID Cross 60.64 IoU). Consequently, the 'just as good' claim is only established for an easy, synthetic subset and cannot be extrapolated to the intended real-world deployment scenario. The authors should either scope the claim explicitly to the synthetic benchmark or provide complementary evidence on unfiltered or real data.

minor comments (4)

[Equations (1)-(2)] The summation notation in the OCR score is garbled ('i=Len(GT)P i=0') and the index range is unclear. Clarify that the sums run over the number of entities and define whether Levenshtein distance is symmetric and computed on normalized strings.
[Table II] The labels 'NBID Cross' and 'MIDV Cross' are easy to misread. A short note in the caption indicating that the training set is the first named dataset and the test set is the other would improve readability.
[Section V, Table III] EasyOCR results are reported but never analyzed; they show much smaller differences across methods (89.80 vs 88.67). A sentence explaining why the two OCR engines behave differently would strengthen the discussion of the metric's sensitivity.
[Section III-B] The photometric augmentation description states probabilities of 5%, 15%, and 100% for negative, blur, and HSV modification, but the last is listed as simply 'modifying the HSV colorspace.' Specify how the HSV modification is applied (e.g., random shifts in H/S/V channels).

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity; only minor self-citation of the authors' own NBID benchmark.

full rationale

The paper's derivation chain is empirical and self-contained. IWPOD-Net is borrowed from prior license-plate detection work, trained from scratch on NBID, and its hyperparameters (data augmentation angles, photometric augmentation on/off) are chosen by validation IoU in Table I. The final comparison in Table III reports test IoU, runtime, and OCR scores computed with external OCR systems (Gemini 2.0, EasyOCR) using a normalized Levenshtein metric defined in Eqs. (1)-(2). No predicted quantity is defined as a fitted parameter, and no outcome is derived from the OCR metric itself. The only self-referential element is the use of the authors' own NBID dataset, including fold splits 'generated by us,' but this is a public benchmark artifact from a prior paper, not an input that forces the efficiency or parity conclusion. The paper also explicitly notes limitations of synthetic data and defers real-data evaluation to future work, which is a validity caveat rather than circularity. Thus the central claims do not reduce by construction to their inputs.

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

The paper's claims depend on the representativeness of the cleaned synthetic dataset, single-run OCR stability, and the transferability of license-plate assumptions. These are not proven inside the paper; they are domain assumptions. The only ad hoc element is the selection of the best augmentation configuration from the ablation table.

free parameters (4)

3D rotation maximum angle sigma = 55 degrees
Selected by validation IoU in Table I; used for all final reported results. It is a training-recipe choice, not a model weight.
Roll-angle cap = 45 degrees
Section IV imposes a roll cap of 45 for sigma > 45; the choice is manual and not justified by data.
Photometric augmentation enabled = false at sigma=55
Ablation Table I shows sigma=55 without photometric gives best validation IoU; the binary augmentation flag is a hand-selected setting.
NBID subset size = 236 images
One of five synthetic instances per real image plus manual exclusion of severe distortions (Section III-A); the central claim is measured only on this subset.

assumptions (4)

domain assumption The NBID subset of 236 images, after manually removing severely distorted documents, is representative of real-world identity-document photos.
Section III-A states the exclusions and relies on NBID's synthetic variety; the conclusion concedes real data remain future work.
domain assumption A single Gemini-2.0 run and the normalized Levenshtein score reliably measure OCR quality differences of about 0.3 points.
Section V reports 'a single run' with no variance or significance testing, yet the SOTA claim depends on small score differences.
domain assumption License-plate affine-warp assumptions transfer to identity documents that may be bent or folded.
Section III-B acknowledges documents can be bent or folded but treats the affine canonical-square representation as sufficient.
ad hoc to paper The selected augmentation configuration (sigma=55, no photometric) chosen by validation IoU is the right configuration to judge the method.
Table I ablation selects this configuration; a different choice could change the headline comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Light-Weight Object Recognition for Real-Time Document Detection." pith.science (2026). https://pith.science/paper/YLSGQBDZ

@misc{pith2026250906246,
  author       = {Pith},
  title        = {Pith review of: Exploring Light-Weight Object Recognition for Real-Time Document Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YLSGQBDZ}},
  note         = {Machine review of arXiv:2509.06246}
}

read the original abstract

Object Recognition and Document Skew Estimation have come a long way in terms of performance and efficiency. New models follow one of two directions: improving performance using larger models, and improving efficiency using smaller models. However, real-time document detection and rectification is a niche that is largely unexplored by the literature, yet it remains a vital step for automatic information retrieval from visual documents. In this work, we strive towards an efficient document detection pipeline that is satisfactory in terms of Optical Character Recognition (OCR) retrieval and faster than other available solutions. We adapt IWPOD-Net, a license plate detection network, and train it for detection on NBID, a synthetic ID card dataset. We experiment with data augmentation and cross-dataset validation with MIDV (another synthetic ID and passport document dataset) to find the optimal scenario for the model. Other methods from both the Object Recognition and Skew Estimation state-of-the-art are evaluated for comparison with our approach. We use each method to detect and rectify the document, which is then read by an OCR system. The OCR output is then evaluated using a novel OCR quality metric based on the Levenshtein distance. Since the end goal is to improve automatic information retrieval, we use the overall OCR quality as a performance metric. We observe that with a promising model, document rectification does not have to be perfect to attain state-of-the-art performance scores. We show that our model is smaller and more efficient than current state-of-the-art solutions while retaining a competitive OCR quality metric. All code is available at https://github.com/BOVIFOCR/iwpod-doc-corners.git

Figures

Figures reproduced from arXiv: 2509.06246 by the authors.

**Figure 2.** Sample document photo from MIDV reach a competitive state-of-the-art result in terms of the final OCR. This model will be included in an information extraction pipeline as the first pre-processing step, aiming at a better information retrieval fidelity at the end (in this case, OCR). As such, we evaluate the model’s effectiveness through the OCR metrics instead of the document detection quality. A. Dataset We use NB… view at source ↗

**Figure 3.** IWPOD-Net Architecture three-step pipeline consisting of a random crop fully containing the object of interest, a 3D rotation with randomized parameters, and a photometric augmentation that alters the color features of the image. These augmentations are done in an online manner, creating new images at training time. The random crop is generated by picking new image width and height parameters, rectifying the licens… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Augmentation Pipeline the final OCR result across different document rectification methods. It is important to note that NBID contains five synthetic instances per real image used for document augmentation. In this work, we use only one of these documents, meaning our …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 16 canonical work pages

[1]

Few shots are all you need: A progressive learning approach for low resource handwritten text recognition,

M. A. Souibgui, A. Forn ´es, Y . Kessentini, and B. Megyesi, “Few shots are all you need: A progressive learning approach for low resource handwritten text recognition,”Pattern Recognition Letters, vol. 160, pp. 43–49, 2022. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S016786552200191X

work page 2022
[2]

Efficient skew detection and correction in scanned document images through clustering of probabilistic hough transforms,

R. Ahmad, S. Naz, and I. Razzak, “Efficient skew detection and correction in scanned document images through clustering of probabilistic hough transforms,”Pattern Recognition Letters, vol. 152, pp. 93–99, 2021. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0167865521003408

work page 2021
[3]

Easyocr,

J. AI, “Easyocr,” 2020. [Online]. Available: https://github.com/jaidedai/ easyocr/

work page 2020
[4]

Nbid dataset: Towards robust information extraction in official documents,

L. Wojcik, L. Coelho, R. Granada, G. F ¨uhr, and D. Menotti, “Nbid dataset: Towards robust information extraction in official documents,” inAnais da XXXVI Conference on Graphics, Patterns and Images. Porto Alegre, RS, Brasil: SBC, 2023, pp. 145–150. [Online]. Available: https://sol.sbc.org.br/index.php/sibgrapi/article/view/27360

work page 2023
[5]

A flexible approach for automatic license plate recognition in unconstrained scenarios,

S. M. Silva and C. R. Jung, “A flexible approach for automatic license plate recognition in unconstrained scenarios,”IEEE Transactions on Intelligent Transportation Systems, 2021

work page 2021
[6]

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

K. Bulatov, E. Emelianova, D. Tropin, N. Skoryukina, Y . Chernyshova, A. Sheshkus, S. Usilin, Z. Ming, J.-C. Burie, M. Luqman, and V . Ar- lazarov, “Midv-2020: a comprehensive benchmark dataset for identity document analysis,”Computer Optics, vol. 46, pp. 252–270, 03 2022

work page 2020
[7]

Gpt-4 technical report,

J. A. et al, “Gpt-4 technical report,” 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257532815

work page 2023
[8]

Gemini: A family of highly capable multimodal models,

G. Team, “Gemini: A family of highly capable multimodal models,”

work page

Show all 22 references

[9]

Benchmarking vision-language models on optical character recognition in dynamic video environments,

S. Nagaonkar, A. Sharma, A. Choithani, and A. Trivedi, “Benchmarking vision-language models on optical character recognition in dynamic video environments,” 02 2025

2025
[10]

Exploring ocr capabilities of gpt-4v(ision) : A quantitative and in-depth evaluation,

Y . Shi, D. Peng, W. Liao, Z. Lin, X. Chen, C. Liu, Y . Zhang, and L. Jin, “Exploring ocr capabilities of gpt-4v(ision) : A quantitative and in-depth evaluation,” 2023. [Online]. Available: https://arxiv.org/abs/2310.16809

2023 arXiv
[11]

doctr: Document text recognition,

Mindee, “doctr: Document text recognition,” https://github.com/mindee/ doctr, 2021

2021
[12]

Character region awareness for text detection,

Y . Baek, B. Lee, D. Han, S. Yun, and H. Lee, “Character region awareness for text detection,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 9365–9374

2019
[13]

Deep residual learning for image recognition,

K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

2016
[14]

An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,

B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 11, pp. 2298–2304, 2017

2017
[15]

You only look once: Unified, real-time object detection,

J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 779–788

2016
[16]

Rtmdet: An empirical study of designing real-time object detectors,

C. Lyu, W. Zhang, H. Huang, Y . Zhou, Y . Wang, Y . Liu, S. Zhang, and K. Chen, “Rtmdet: An empirical study of designing real-time object detectors,”ArXiv, vol. abs/2212.07784, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:254685870

2022 arXiv
[17]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2999–3007

2017
[18]

Feature pyramid networks for object detection,

T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” in2017 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 936–944

2017
[19]

Adaptive radial projection on fourier magnitude spectrum for document image skew estimation,

T. M. Luan Pham, Hao Hoang and T. A. Tran, “Adaptive radial projection on fourier magnitude spectrum for document image skew estimation,” in2022 29th International Conference on Image Processing (ICIP). IEEE, 2022

2022
[20]

License plate detection and recognition in unconstrained scenarios,

S. M. Silva and C. R. Jung, “License plate detection and recognition in unconstrained scenarios,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 580–596

2018
[21]

Yolov11: An overview of the key architectural enhancements,

R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,” 10 2024

2024
[2024]

Available: https://arxiv.org/abs/2312.11805

[Online]. Available: https://arxiv.org/abs/2312.11805

arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Few shots are all you need: A progressive learning approach for low resource handwritten text recognition,

M. A. Souibgui, A. Forn ´es, Y . Kessentini, and B. Megyesi, “Few shots are all you need: A progressive learning approach for low resource handwritten text recognition,”Pattern Recognition Letters, vol. 160, pp. 43–49, 2022. [Online]. Available: https: //www.sciencedirect.com/science/article/pii/S016786552200191X

work page 2022

[2] [2]

Efficient skew detection and correction in scanned document images through clustering of probabilistic hough transforms,

R. Ahmad, S. Naz, and I. Razzak, “Efficient skew detection and correction in scanned document images through clustering of probabilistic hough transforms,”Pattern Recognition Letters, vol. 152, pp. 93–99, 2021. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S0167865521003408

work page 2021

[3] [3]

Easyocr,

J. AI, “Easyocr,” 2020. [Online]. Available: https://github.com/jaidedai/ easyocr/

work page 2020

[4] [4]

Nbid dataset: Towards robust information extraction in official documents,

L. Wojcik, L. Coelho, R. Granada, G. F ¨uhr, and D. Menotti, “Nbid dataset: Towards robust information extraction in official documents,” inAnais da XXXVI Conference on Graphics, Patterns and Images. Porto Alegre, RS, Brasil: SBC, 2023, pp. 145–150. [Online]. Available: https://sol.sbc.org.br/index.php/sibgrapi/article/view/27360

work page 2023

[5] [5]

A flexible approach for automatic license plate recognition in unconstrained scenarios,

S. M. Silva and C. R. Jung, “A flexible approach for automatic license plate recognition in unconstrained scenarios,”IEEE Transactions on Intelligent Transportation Systems, 2021

work page 2021

[6] [6]

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

K. Bulatov, E. Emelianova, D. Tropin, N. Skoryukina, Y . Chernyshova, A. Sheshkus, S. Usilin, Z. Ming, J.-C. Burie, M. Luqman, and V . Ar- lazarov, “Midv-2020: a comprehensive benchmark dataset for identity document analysis,”Computer Optics, vol. 46, pp. 252–270, 03 2022

work page 2020

[7] [7]

Gpt-4 technical report,

J. A. et al, “Gpt-4 technical report,” 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:257532815

work page 2023

[8] [8]

Gemini: A family of highly capable multimodal models,

G. Team, “Gemini: A family of highly capable multimodal models,”

work page