pith. sign in

arxiv: 2604.08741 · v1 · submitted 2026-04-09 · 💻 cs.CV

LPLCv2: An Expanded Dataset for Fine-Grained License Plate Legibility Classification

Lucas Wojcik , Eduardo A. F. Machoski , Eduil Nascimento Jr. , Rayson Laroca , David Menotti This is my paper

Pith reviewed 2026-05-10 17:18 UTC · model grok-4.3

classification 💻 cs.CV
keywords license plate recognitionlegibility classificationALPRdatasetannotationfine-grained classificationcomputer visionF1-score
0
0 comments X

The pith

Expanding and refining a license plate legibility dataset with new multi-level labels and training adjustments raises baseline F1-score to 89.5%.

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

The paper expands an existing small benchmark for classifying whether license plates are legible in real-world photos. It more than triples the data volume by adding capture days, corrects prior annotation mistakes, and layers in new labels: bounding boxes and text at the plate level, vehicle make model type and color at the vehicle level, plus camera identity, weather, and timing at the image level. A training procedure is introduced that uses an exponential moving average loss function together with a tuned learning rate scheduler. These steps let a standard model reach 89.5 percent F1 on the test set, well above earlier results. A separate split protocol is proposed to limit camera overlap between training and evaluation, with results indicating only modest impact from such contamination.

Core claim

We present LPLCv2, an expanded dataset for fine-grained license plate legibility classification that is more than three times larger than its predecessor, with revised annotations and novel labels at the license plate, vehicle, and image levels. A baseline model trained with an exponential moving average loss function and refined learning rate scheduler achieves an 89.5% F1-score on the test set, substantially exceeding the prior state of the art. The work also defines a protocol that explicitly controls for camera contamination across train and test splits, where experiments indicate limited effect on performance.

What carries the argument

The LPLCv2 dataset with its plate-level, vehicle-level, and image-level annotations together with the exponential moving average loss function and refined learning rate scheduler.

If this is right

  • ALPR systems can more reliably flag illegible plates for manual review or recapture in uncontrolled environments.
  • The added vehicle and condition labels support analysis of which factors most degrade legibility.
  • The camera-contamination protocol offers a reusable method for constructing fair evaluation splits in other imaging datasets.
  • Public release of the full dataset and code enables direct replication and extension by other researchers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The multi-level labels could support auxiliary prediction tasks that improve overall automatic license plate recognition pipelines even when plates are only partially legible.
  • The training adjustments may generalize to other fine-grained classification problems involving compressed or low-quality imagery.
  • Deployed systems might combine legibility scores with vehicle metadata to trigger targeted re-imaging or alternative recognition routes.

Load-bearing premise

The revised annotations are accurate and the added vehicle-level and image-level labels meaningfully improve legibility classification without introducing new biases or annotation errors.

What would settle it

An independent audit of a random sample of test-set annotations that finds substantial disagreement with the published labels, or an ablation experiment in which the new labels and exponential moving average training steps are removed and F1-score falls to or below the previous state of the art.

Figures

Figures reproduced from arXiv: 2604.08741 by David Menotti, Eduardo A. F. Machoski, Eduil Nascimento Jr., Lucas Wojcik, Rayson Laroca.

Figure 1
Figure 1. Figure 1: LPs grouped by legibility class, following the definition of [ [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative samples from the proposed LPLCv2 dataset. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: LP legibility annotation errors observed in LPLCv1 (left) and their [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Confusion matrix from one of the EMA experiment folds. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of misclassified LPs illustrating borderline legibility cases. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LPs misclassified into non-adjacent legibility classes. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

Modern Automatic License Plate Recognition (ALPR) systems achieve outstanding performance in controlled, well-defined scenarios. However, large-scale real-world usage remains challenging due to low-quality imaging devices, compression artifacts, and suboptimal camera installation. Identifying illegible license plates (LPs) has recently become feasible through a dedicated benchmark; however, its impact has been limited by its small size and annotation errors. In this work, we expand the original benchmark to over three times the size with two extra capture days, revise its annotations and introduce novel labels. LP-level annotations include bounding boxes, text, and legibility level, while vehicle-level annotations comprise make, model, type, and color. Image-level annotations feature camera identity, capture conditions (e.g., rain and faulty cameras), acquisition time, and day ID. We present a novel training procedure featuring an Exponential Moving Average-based loss function and a refined learning rate scheduler, addressing common mistakes in testing. These improvements enable a baseline model to achieve an 89.5% F1-score on the test set, considerably surpassing the previous state of the art. We further introduce a novel protocol to explicitly addresses camera contamination between training and evaluation splits, where results show a small impact. Dataset and code are publicly available at https://github.com/lmlwojcik/LPLCv2-Dataset.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript presents LPLCv2, an expanded and revised dataset for license plate legibility classification. It expands the original benchmark by more than three times with additional data from two capture days, revises annotations, and adds multi-level labels: LP-level (bounding boxes, text, legibility), vehicle-level (make, model, type, color), and image-level (camera ID, conditions, time, day). A novel training procedure with an EMA-based loss and refined LR scheduler is introduced to mitigate testing errors. This enables a baseline model to reach 89.5% F1-score on the test set, exceeding prior SOTA. A camera contamination protocol is also proposed and evaluated, showing minimal impact. The dataset and code are made publicly available.

Significance. Should the central performance claims be substantiated through additional validation of the annotation process, this contribution would offer a valuable, larger-scale resource for developing robust ALPR systems capable of handling real-world degradations. The multi-faceted annotations enable richer analysis and potential multi-task learning. Public release of data and code supports reproducibility. The camera protocol addresses an important split contamination issue. However, the current evidence for the gains being due to the proposed methods rather than label revisions is insufficient.

major comments (3)
  1. §3.2 Annotation Revision: The paper reports revising annotations but provides no quantitative analysis of changes (e.g., percentage of labels altered) or inter-annotator agreement metrics. This information is critical to determine if the reported F1-score improvement of 89.5% is attributable to the dataset expansion and training procedure or primarily to corrected labels.
  2. §5 Experiments and Results: Missing ablation studies comparing performance of the baseline model on the original dataset labels versus the revised LPLCv2 annotations. Without this, the individual contributions of the EMA loss, LR scheduler, and label revisions cannot be disentangled, weakening support for the novel training procedure.
  3. §4.1 Novel Training Procedure: The EMA-based loss function and refined learning rate scheduler are claimed to address common testing mistakes, but the manuscript lacks a detailed error analysis or examples of mistakes corrected by these components, making their effectiveness hard to evaluate.
minor comments (2)
  1. Abstract: The previous state-of-the-art F1-score should be explicitly stated alongside the new 89.5% result for immediate context.
  2. Related Work: Ensure all prior ALPR legibility works are cited, particularly the original LPLC benchmark paper.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and indicate the revisions we will make to strengthen the paper.

read point-by-point responses

  1. Referee: §3.2 Annotation Revision: The paper reports revising annotations but provides no quantitative analysis of changes (e.g., percentage of labels altered) or inter-annotator agreement metrics. This information is critical to determine if the reported F1-score improvement of 89.5% is attributable to the dataset expansion and training procedure or primarily to corrected labels.

    Authors: We agree that a quantitative analysis of the revisions would improve transparency. In the revised manuscript, we will add a table in Section 3.2 reporting the percentage of labels altered across categories (legibility, text, bounding boxes, and vehicle attributes). The revisions were performed by a single expert following a documented protocol to correct clear errors from the original dataset rather than as a fresh multi-annotator labeling effort; therefore inter-annotator agreement was not computed. We will expand the section to describe the revision protocol in detail. revision: partial

  2. Referee: §5 Experiments and Results: Missing ablation studies comparing performance of the baseline model on the original dataset labels versus the revised LPLCv2 annotations. Without this, the individual contributions of the EMA loss, LR scheduler, and label revisions cannot be disentangled, weakening support for the novel training procedure.

    Authors: We acknowledge that such ablations would help isolate contributions. Because the test set annotations were also revised, we will add experiments in the revised Section 5 that (i) train and evaluate the baseline using the original LPLC labels where they exist and (ii) compare against the new training procedure on the revised LPLCv2 splits. These results will clarify the relative impact of label corrections versus the proposed EMA loss and scheduler. revision: yes

  3. Referee: §4.1 Novel Training Procedure: The EMA-based loss function and refined learning rate scheduler are claimed to address common testing mistakes, but the manuscript lacks a detailed error analysis or examples of mistakes corrected by these components, making their effectiveness hard to evaluate.

    Authors: We will revise Section 4.1 to include a dedicated error analysis subsection. This will catalog the most frequent testing mistakes (e.g., false negatives on low-contrast plates under rain or compression) and provide qualitative examples showing how the EMA loss and adjusted scheduler reduce these errors relative to standard cross-entropy training. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical dataset expansion and baseline results are self-contained

full rationale

The paper describes dataset expansion (>3x size), annotation revisions, new multi-level labels, a training procedure with EMA loss and LR scheduler, and reports 89.5% F1 on a held-out test set. No equations, derivations, or first-principles claims appear that could reduce to fitted inputs by construction. No self-citations are invoked to justify uniqueness theorems or ansatzes. Performance gains are presented as direct outcomes of the new data and training choices rather than statistical artifacts of parameter fitting or renaming. The work is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The work relies on standard assumptions in computer vision dataset curation and supervised classification; no free parameters, axioms, or invented entities are introduced beyond conventional practices.

pith-pipeline@v0.9.0 · 5555 in / 1034 out tokens · 56075 ms · 2026-05-10T17:18:58.498159+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    Auto- matic license plate recognition in in-the-wild scenarios: A comprehen- sive review, open issues, and future directions,

    A. Ismail, M. Mehri, A. Sahbani, and N. Essoukri Ben Amara, “Auto- matic license plate recognition in in-the-wild scenarios: A comprehen- sive review, open issues, and future directions,”IEEE Access, vol. 13, pp. 145 387–145 415, 2025

    work page 2025

  2. [2]

    Advancing multinational license plate recognition through synthetic and real data fusion: A comprehensive evaluation,

    R. Laroca, V . Estevam, G. J. P. Moreira, R. Minetto, and D. Menotti, “Advancing multinational license plate recognition through synthetic and real data fusion: A comprehensive evaluation,”IET Intelligent Transport Systems, vol. 19, no. 1, p. e70086, 2025

    work page 2025

  3. [3]

    An ultra-fast automatic license plate recognition approach for unconstrained scenarios,

    X. Ke, G. Zeng, and W. Guo, “An ultra-fast automatic license plate recognition approach for unconstrained scenarios,”IEEE Transactions on Intelligent Transportation Systems, vol. 24, pp. 5172–5185, 2023

    work page 2023

  4. [4]

    An end-to-end contrastive license plate detector,

    H. Ding, J. Gao, Y . Yuan, and Q. Wang, “An end-to-end contrastive license plate detector,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 1, pp. 503–516, 2024

    work page 2024

  5. [5]

    Benchmark for license plate character segmentation,

    G. R. Gonc ¸alves, S. P. G. da Silva, D. Menotti, and W. R. Schwartz, “Benchmark for license plate character segmentation,”Journal of Elec- tronic Imaging, vol. 25, no. 5, p. 053034, 2016

    work page 2016

  6. [6]

    EnglishLP Database,

    V . Srebri ´c, “EnglishLP Database,” 2003. [Online]. Available: https: //www.zemris.fer.hr/projects/LicensePlates/english/baza slika.zip

    work page 2003

  7. [7]

    A robust attentional framework for license plate recognition in the wild,

    L. Zhanget al., “A robust attentional framework for license plate recognition in the wild,”IEEE Transactions on Intelligent Transportation Systems, vol. 22, no. 11, pp. 6967–6976, 2021

    work page 2021

  8. [8]

    Fog and rain augmentation for license plate recognition in tropical country environment,

    S. Wahyuet al., “Fog and rain augmentation for license plate recognition in tropical country environment,”IAES International Journal of Artificial Intelligence (IJ-AI), vol. 13, no. 4, p. 3951, Dec. 2024

    work page 2024

  9. [9]

    Toward advancing license plate super-resolution in real-world scenarios: A dataset and benchmark,

    V . Nascimento, G. E. Lima, R. O. Ribeiro, W. R. Schwartz, R. Laroca, and D. Menotti, “Toward advancing license plate super-resolution in real-world scenarios: A dataset and benchmark,”Journal of the Brazilian Computer Society, vol. 1, no. 31, pp. 435–449, 2025

    work page 2025

  10. [10]

    LPLC: A dataset for license plate legibility classification,

    L. Wojcik, G. E. Lima, V . Nascimento, E. Nascimento Jr., R. Laroca, and D. Menotti, “LPLC: A dataset for license plate legibility classification,” Conference on Graphics, Patterns and Images (SIBGRAPI), 2025

    work page 2025

  11. [11]

    Enhancing license plate super-resolution: A layout-aware and character-driven approach,

    V . Nascimento, R. Laroca, R. O. Ribeiro, W. R. Schwartz, and D. Menotti, “Enhancing license plate super-resolution: A layout-aware and character-driven approach,”Conference on Graphics, Patterns and Images (SIBGRAPI), pp. 1–6, 2024

    work page 2024

  12. [12]

    LPSRGAN: Generative adversarial networks for super-resolution of license plate image,

    Y . Pan, J. Tang, and T. Tjahjadi, “LPSRGAN: Generative adversarial networks for super-resolution of license plate image,”Neurocomputing, vol. 580, p. 127426, 2024

    work page 2024

  13. [13]

    Leveraging model fusion for improved license plate recognition,

    R. Laroca, L. A. Zanlorensi, V . Estevam, R. Minetto, and D. Menotti, “Leveraging model fusion for improved license plate recognition,” in Iberoamerican Congress on Pattern Recognition, Nov 2023, pp. 60–75

    work page 2023

  14. [14]

    Improving multi-type license plate recognition via learning globally and contrastively,

    Q. Liuet al., “Improving multi-type license plate recognition via learning globally and contrastively,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 9, pp. 11 092–11 102, 2024

    work page 2024

  15. [15]

    A robust real-time automatic license plate recognition based on the YOLO detector,

    R. Laroca, E. Severo, L. A. Zanlorensi, L. S. Oliveira, G. R. Gonc ¸alves, W. R. Schwartz, and D. Menotti, “A robust real-time automatic license plate recognition based on the YOLO detector,” inInternational Joint Conference on Neural Networks (IJCNN), July 2018, pp. 1–10

    work page 2018

  16. [16]

    Real-time license plate detection and recognition using deep convolutional neural networks,

    S. M. Silva and C. R. Jung, “Real-time license plate detection and recognition using deep convolutional neural networks,”Journal of Visual Communication and Image Representation, vol. 71, p. 102773, 2020

    work page 2020

  17. [17]

    License plate recognition using a hybrid class attention- inception network,

    A. Alzahrani, “License plate recognition using a hybrid class attention- inception network,” inInternational Conference on Modelling Strategies in Mathematics, vol. 3306, 2025, p. 060023

    work page 2025

  18. [18]

    A robust layout-independent license plate detection and recognition model based on attention method,

    T.-M. Seo and D.-J. Kang, “A robust layout-independent license plate detection and recognition model based on attention method,”IEEE Access, vol. 10, pp. 57 427–57 436, 2022

    work page 2022

  19. [19]

    Irregular license plate recognition via global information integration,

    Y .-Y . Liu, Q. Liu, S.-L. Chen, F. Chen, and X.-C. Yin, “Irregular license plate recognition via global information integration,” inInternational Conference on Multimedia Modeling, 2024, pp. 325–339

    work page 2024

  20. [20]

    On the cross-dataset generalization in license plate recognition,

    R. Laroca, E. V . Cardoso, D. R. Lucio, V . Estevam, and D. Menotti, “On the cross-dataset generalization in license plate recognition,” in International Conference on Computer Vision Theory and Applications (VISAPP), Feb 2022, pp. 166–178

    work page 2022

  21. [21]

    Towards end-to-end license plate detection and recognition: A large dataset and baseline,

    Z. Xu, W. Yang, A. Meng, N. Lu, H. Huang, C. Ying, and L. Huang, “Towards end-to-end license plate detection and recognition: A large dataset and baseline,” inEuropean Conference on Computer Vision (ECCV), 2018, pp. 261–277

    work page 2018

  22. [22]

    GroupPlate: Toward multi-category license plate recognition,

    Y . Gao, H. Lu, S. Mu, and S. Xu, “GroupPlate: Toward multi-category license plate recognition,”IEEE Transactions on Intelligent Transporta- tion Systems, vol. 24, no. 5, pp. 5586–5599, 2023

    work page 2023

  23. [23]

    Do we train on test data? The impact of near-duplicates on license plate recognition,

    R. Laroca, V . Estevam, A. S. Britto Jr., R. Minetto, and D. Menotti, “Do we train on test data? The impact of near-duplicates on license plate recognition,” inInternational Joint Conference on Neural Networks (IJCNN), June 2023, pp. 1–8

    work page 2023

  24. [24]

    Application-oriented license plate recognition,

    G. S. Hsu, J. C. Chen, and Y . Z. Chung, “Application-oriented license plate recognition,”IEEE Transactions on Vehicular Technology, vol. 62, no. 2, pp. 552–561, Feb 2013

    work page 2013

  25. [25]

    License plate detection and recognition in unconstrained scenarios,

    S. M. Silva and C. R. Jung, “License plate detection and recognition in unconstrained scenarios,” inEuropean Conference on Computer Vision (ECCV), Sept 2018, pp. 593–609

    work page 2018

  26. [26]

    OpenALPR-EU Dataset,

    OpenALPR, “OpenALPR-EU Dataset,” 2016. [Online]. Available: https://github.com/openalpr/benchmarks/tree/master/endtoend/eu

    work page 2016

  27. [27]

    YOLOv11,

    Ultralytics, “YOLOv11,” 2025, accessed: 2026-01-28. [Online]. Available: https://docs.ultralytics.com/models/yolo11/

    work page 2025

  28. [28]

    Scene text recognition with permuted autoregressive sequence models,

    D. Bautista and R. Atienza, “Scene text recognition with permuted autoregressive sequence models,” inEuropean Conference on Computer Vision (ECCV), 2022, pp. 178–196

    work page 2022

  29. [29]

    Deep residual learning for image recognition,

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

    work page 2016

  30. [30]

    A first look at dataset bias in license plate recognition,

    R. Laroca, M. Santos, V . Estevam, E. Luz, and D. Menotti, “A first look at dataset bias in license plate recognition,” inConference on Graphics, Patterns and Images (SIBGRAPI), Oct 2022, pp. 234–239

    work page 2022

  31. [31]

    Exponential Moving Average of Weights in Deep Learning: Dynamics and Benefits,

    D. Morales Brotons, T. V ogels, and H. Hendrikx, “Exponential Moving Average of Weights in Deep Learning: Dynamics and Benefits,”Trans- actions on Machine Learning Research Journal, pp. 1–27, Apr. 2024

    work page 2024

  32. [32]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR), 2015

    work page 2015

  33. [33]

    Toward enhancing vehicle color recognition in ad- verse conditions: A dataset and benchmark,

    G. E. Limaet al., “Toward enhancing vehicle color recognition in ad- verse conditions: A dataset and benchmark,” inConference on Graphics, Patterns and Images (SIBGRAPI), Sept 2024, pp. 1–6

    work page 2024