pith. sign in

arxiv: 2606.05018 · v1 · pith:2K2HTCAEnew · submitted 2026-06-03 · 💻 cs.CV

Handwriting Extraction and Analysis of Signature Lists in Swiss Popular Initiatives

Marco Peer , Thomas Gorges , Mathias Seuret , Vincent Christlein , Andreas Fischer This is my paper

Pith reviewed 2026-06-28 06:07 UTC · model grok-4.3

classification 💻 cs.CV
keywords handwriting analysiswriter retrievalsignature validationduplicate detectionOCRSwiss popular initiativesdocument analysis
0
0 comments X

The pith

Writer retrieval methods can identify similar handwriting entries to support duplicate detection in Swiss signature lists.

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

The paper examines automated tools to ease the manual validation of handwritten signatures for Swiss popular initiatives and referendums. It tests a pipeline of line segmentation, text recognition, and writer retrieval on 443 entries from 418 writers. OCR produces high error rates on short names and addresses, but writer retrieval reaches 50.6 percent mAP in matching visually similar entries. This positions writer retrieval as a practical aid for flagging potential duplicate submissions based on handwriting.

Core claim

A pipeline of template-based line segmentation, OCR, and writer retrieval was tested on 443 handwritten entries from 418 writers. OCR reaches a character error rate of 29.6 percent on first names and struggles with out-of-vocabulary terms, while writer retrieval achieves 50.6 percent mAP and more reliably identifies visually similar entries across lists to support duplicate detection.

What carries the argument

Writer retrieval techniques that match entries by handwriting similarity across signature lists.

If this is right

  • Writer retrieval can flag potential duplicate submissions for manual review.
  • Off-the-shelf OCR systems remain unreliable for transcribing short handwritten names and addresses.
  • The segmentation-plus-retrieval pipeline offers a concrete way to assist labor-intensive list validation.
  • Performance metrics are measured on a dataset built from real signature entries.

Where Pith is reading between the lines

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

  • Such methods could be integrated into existing validation workflows to reduce the volume of entries requiring full manual inspection.
  • Collecting additional Swiss handwriting samples might raise the mAP beyond the reported 50.6 percent.
  • The same retrieval approach could extend to other administrative tasks that require spotting repeated handwritten content.

Load-bearing premise

The 443-entry dataset collected from 418 writers represents the handwriting variability, formats, and duplicate patterns found in real Swiss signature lists.

What would settle it

Apply the writer retrieval method to a new collection of actual submitted signature lists from multiple initiatives and check whether the top-ranked similar pairs are genuine duplicates or forgeries.

Figures

Figures reproduced from arXiv: 2606.05018 by Andreas Fischer, Marco Peer, Mathias Seuret, Thomas Gorges, Vincent Christlein.

Figure 1
Figure 1. Figure 1: Overview of the methodology proposed. The handwriting on the form is blurred to preserve privacy. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of the templates. Template Identification: Layout-based features are extracted using Histogram of Oriented Gradients (HOG) [17]. Before feature extraction, scans are grouped by orientation and nearly empty or corrupted pages are removed. These features are em￾bedded with UMAP [18] and clustered using HDBSCAN [19], which handles an unknown number of templates and identifies outliers as noise. To re… view at source ↗
Figure 3
Figure 3. Figure 3: Writer Retrieval. Documents written by writer A should [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE visualization of the writers showing that most [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
read the original abstract

Popular initiatives and referendums are central to Swiss democracy, yet the validation of handwritten signature lists remains a labor-intensive manual process. This paper investigates the potential of automated document analysis methods, including OCR and AI-based handwriting analysis, to support this task. We propose a pipeline combining template-based line segmentation with text recognition and writer retrieval techniques, evaluated on a dataset of 443 handwritten entries from 418 writers. Results show that OCR struggles with out-of-vocabulary handwriting, with a CER of 29.6% for first names. In contrast, writer retrieval performs more robustly, reaching an mAP of 50.6%. Furthermore, our experiments indicate that off-the-shelf OCR systems are not sufficiently reliable for transcription of handwritten signature data, particularly for short, out-of-vocabulary entries such as names or addresses. However, writer retrieval methods can effectively identify visually similar entries across signature lists, making them a suitable tool for supporting the detection of potential duplicate submissions based on handwriting similarity.

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

2 major / 1 minor

Summary. The paper proposes a pipeline combining template-based line segmentation, OCR-based text recognition, and writer retrieval to support validation of handwritten signature lists for Swiss popular initiatives. It evaluates the approach on an internal dataset of 443 entries from 418 writers, reporting a CER of 29.6% for first-name transcription and an mAP of 50.6% for writer retrieval. The central claim is that while off-the-shelf OCR is unreliable for short out-of-vocabulary entries, writer retrieval methods can effectively identify visually similar entries and are thus suitable for supporting detection of potential duplicate submissions based on handwriting similarity.

Significance. If the reported mAP generalizes, the work provides a concrete baseline (CER 29.6%, mAP 50.6%) for applying document analysis techniques to a real civic process, potentially reducing manual validation effort. The explicit reporting of these metrics on a non-trivial collection of signature data is a strength that future studies can build upon.

major comments (2)
  1. [Dataset] Dataset section: The collection comprises 443 entries from 418 writers (implying ~25 repeated writers). The evaluation therefore rests on a small number of near-duplicates whose similarity patterns may not reflect operational cases such as forgeries, aging signatures, or cross-list variations. This directly limits the strength of the claim that the 50.6% mAP demonstrates suitability for supporting duplicate detection in practice.
  2. [Experiments] Experiments / Evaluation protocol: No external validation set drawn from official Swiss submissions is described, nor is the precise construction of queries versus gallery (or definition of a positive duplicate match) detailed. Without these, it is unclear whether the mAP measures the operational task the authors conclude it supports.
minor comments (1)
  1. [Abstract] Abstract: The evaluation protocol, baselines, and error analysis are not summarized, making the reported numbers difficult to interpret at first reading.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our dataset and evaluation protocol. The comments correctly identify limitations in scale and external validation that affect the strength of our claims. We address each point below and will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Dataset] Dataset section: The collection comprises 443 entries from 418 writers (implying ~25 repeated writers). The evaluation therefore rests on a small number of near-duplicates whose similarity patterns may not reflect operational cases such as forgeries, aging signatures, or cross-list variations. This directly limits the strength of the claim that the 50.6% mAP demonstrates suitability for supporting duplicate detection in practice.

    Authors: We agree that the limited number of repeated writers (~25) restricts evaluation on challenging variations such as forgeries or aging. Our internal collection was assembled to mirror the low duplicate rate typical of real signature lists. The mAP of 50.6% shows the retrieval method can surface the existing duplicates present in this data. In revision we will add explicit discussion of these constraints in the limitations section and moderate the language on operational suitability. revision: partial

  2. Referee: [Experiments] Experiments / Evaluation protocol: No external validation set drawn from official Swiss submissions is described, nor is the precise construction of queries versus gallery (or definition of a positive duplicate match) detailed. Without these, it is unclear whether the mAP measures the operational task the authors conclude it supports.

    Authors: Official Swiss submission data cannot be used due to privacy restrictions; the reported results are therefore based on our internal collection. We will revise the Experiments section to specify the query-gallery split, the exact definition of a positive match (entries from the same writer), and how mAP is computed. This will clarify the relation between the reported metric and the duplicate-flagging use case. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation with no derivations or self-referential reductions

full rationale

The paper collects a dataset of 443 entries, applies existing OCR and writer retrieval methods, and reports direct performance metrics (CER 29.6%, mAP 50.6%). No equations, fitted parameters, predictions, or derivations are present that could reduce to inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems or ansatzes. The central claim is an empirical observation about method suitability on the internal data, which does not involve any circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Empirical application study; no free parameters, axioms, or invented entities introduced.

pith-pipeline@v0.9.1-grok · 5705 in / 1120 out tokens · 27897 ms · 2026-06-28T06:07:38.916639+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    Trocr: Transformer-based optical character recognition with pre-trained models,

    M. Li, T. Lv, J. Chen, L. Cui, Y . Lu, D. A. F. Flor ˆencio, C. Zhang, Z. Li, and F. Wei, “Trocr: Transformer-based optical character recognition with pre-trained models,” inThirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, 2023, pp. 13 094–13 102

    2023

  2. [2]

    Writer retrieval for historical documents,

    M. Peer, “Writer retrieval for historical documents,” Ph.D. dissertation, 2025

    2025

  3. [3]

    Handwriting Analysis with Focus on Writer Identification and Writer Retrieval,

    V . Christlein, “Handwriting Analysis with Focus on Writer Identification and Writer Retrieval,” Ph.D. dissertation, 2018

    2018

  4. [4]

    Layoutlmv3: Pre-training for document AI with unified text and image masking,

    Y . Huang, T. Lv, L. Cui, Y . Lu, and F. Wei, “Layoutlmv3: Pre-training for document AI with unified text and image masking,” inMM ’22: The 30th ACM International Conference on Multimedia 2022, 2022, pp. 4083–4091

    2022

  5. [5]

    Ocr-free document understanding transformer,

    G. Kim, T. Hong, M. Yim, J. Nam, J. Park, J. Yim, W. Hwang, S. Yun, D. Han, and S. Park, “Ocr-free document understanding transformer,” in Computer Vision - ECCV 2022 - 17th European Conference, 2022, pp. 498–517

    2022

  6. [6]

    A novel connectionist system for unconstrained handwriting recognition,

    A. Graves, M. Liwicki, S. Fern ´andez, R. Bertolami, H. Bunke, and J. Schmidhuber, “A novel connectionist system for unconstrained handwriting recognition,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 5, pp. 855–868, 2009. TABLE III: Results of different writer retrieval methods and datasets showing that a supervised use of additional training da...

    2009

  7. [7]

    Handwritten text recognition: A survey,

    C. Garrido-Munoz, A. Rios-Vila, and J. Calvo-Zaragoza, “Handwritten text recognition: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 4, p. 4367–4387, Apr. 2026

    2026

  8. [8]

    Qwen2.5-Coder Technical Report

    Qwen Team, “Qwen3 technical report,”arXiv preprint, 2024, arXiv:2409.12186

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    Benchmarking large language models for handwritten text recognition,

    G. Crosilla, L. Klic, and G. Colavizza, “Benchmarking large language models for handwritten text recognition,”Journal of Documentation, vol. 81, no. 7, pp. 334–354, 2025

    2025

  10. [10]

    Writer identification using vlad encoded contour-zernike moments,

    V . Christlein, D. Bernecker, and E. Angelopoulou, “Writer identification using vlad encoded contour-zernike moments,” in2015 13th International Conference on Document Analysis and Recognition (ICDAR), 2015, pp. 906–910

    2015

  11. [11]

    Writer retrieval using compact con- volutional transformers and netmvlad,

    M. Peer, F. Kleber, and R. Sablatnig, “Writer retrieval using compact con- volutional transformers and netmvlad,” in26th International Conference on Pattern Recognition, ICPR 2022, 2022, pp. 1571–1578

    2022

  12. [12]

    Learning features for writer retrieval and identification using triplet cnns,

    M. Keglevic, S. Fiel, and R. Sablatnig, “Learning features for writer retrieval and identification using triplet cnns,” in16th International Conference on Frontiers in Handwriting Recognition, ICFHR 2018, 2018, pp. 211–216

    2018

  13. [13]

    SAGHOG: Self-Supervised Autoencoder for Generating HOG Features for Writer Retrieval,

    M. Peer, F. Kleber, and R. Sablatnig, “SAGHOG: Self-Supervised Autoencoder for Generating HOG Features for Writer Retrieval,” in Document Analysis and Recognition - ICDAR 2024, 2024, pp. 121–138

    2024

  14. [14]

    Self-supervised vision transformers for writer retrieval,

    T. Raven, A. Matei, and G. A. Fink, “Self-supervised vision transformers for writer retrieval,” inDocument Analysis and Recognition - ICDAR 2024, E. H. Barney Smith, M. Liwicki, and L. Peng, Eds. Cham: Springer Nature Switzerland, 2024, pp. 380–396

    2024

  15. [15]

    Unsupervised feature learning for writer identification and writer retrieval,

    V . Christlein, M. Gropp, S. Fiel, and A. K. Maier, “Unsupervised feature learning for writer identification and writer retrieval,” in14th IAPR International Conference on Document Analysis and Recognition, ICDAR 2017,, 2017, pp. 991–997

    2017

  16. [16]

    Questioned document examination using cedar-fox,

    S. N. Srihari, B. Srinivasan, and K. Desai, “Questioned document examination using cedar-fox,”Journal of Forensic Document Examination, vol. 28, p. 15–26, Dec. 2018

    2018

  17. [17]

    Histograms of oriented gradients for human detection,

    N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” inProceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), 2005, pp. 886–893

    2005

  18. [18]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction

    L. McInnes, J. Healy, and J. Melville, “Umap: Uniform manifold approximation and projection for dimension reduction,”arXiv preprint arXiv:1802.03426, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  19. [19]

    Hdbscan: Hierarchical density based clustering,

    L. McInnes, J. Healy, and S. Astels, “Hdbscan: Hierarchical density based clustering,”Journal of Open Source Software, vol. 2, no. 11, p. 205, 2017

    2017

  20. [20]

    Distinctive image features from scale-invariant keypoints,

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004

    2004

  21. [21]

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography,”Communications of the ACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  22. [22]

    Parametric image alignment using enhanced correlation coefficient maximization,

    G. D. Evangelidis and E. Z. Psarakis, “Parametric image alignment using enhanced correlation coefficient maximization,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 30, no. 10, pp. 1858–1865, 2008

    2008

  23. [23]

    A computational approach to edge detection,

    J. Canny, “A computational approach to edge detection,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. PAMI-8, no. 6, pp. 679–698, 1986

    1986

  24. [24]

    Method and means for recognizing complex patterns,

    P. V . C. Hough, “Method and means for recognizing complex patterns,” U.S. Patent, no. 3,069,654, 1962

    1962

  25. [25]

    The iam-database: An english sentence database for offline handwriting recognition,

    U.-V . Marti and H. Bunke, “The iam-database: An english sentence database for offline handwriting recognition,”International Journal on Document Analysis and Recognition, vol. 5, no. 1, pp. 39–46, 2002

    2002

  26. [26]

    CVL-DataBase: An off- line database for writer retrieval, writer identification and word spotting,

    F. Kleber, S. Fiel, M. Diem, and R. Sablatnig, “CVL-DataBase: An off- line database for writer retrieval, writer identification and word spotting,” in12th International Conference on Document Analysis and Recognition, ICDAR 2013, 2013, pp. 560–564

    2013

  27. [27]

    Hybrid page layout analysis via tab-stop detection,

    R. Smith, “Hybrid page layout analysis via tab-stop detection,”Proceed- ings of the Tenth International Conference on Document Analysis and Recognition (ICDAR), pp. 241–245, 2009

    2009

  28. [28]

    Kraken: An universal text recognizer for the humanities,

    B. Kiessling, C. Reul, M. Wehner, and U. Springmann, “Kraken: An universal text recognizer for the humanities,”Proceedings of the Digital Humanities Conference, 2019

    2019

  29. [29]

    arXiv preprint arXiv:2009.09941 (2020)

    Y . Du, C. Li, R. Guo, X. Yin, W. Liu, J. Zhouet al., “Pp-ocr: A practical ultra lightweight ocr system,”arXiv preprint, 2020, arXiv:2009.09941

    work page arXiv 2020

  30. [30]

    Towards the influence of text quantity on writer retrieval,

    M. Peer, R. Sablatnig, and F. Kleber, “Towards the influence of text quantity on writer retrieval,” inDocument Analysis and Recognition – ICDAR 2025, 2026, pp. 129–145

    2025

  31. [31]

    All About VLAD,

    R. Arandjelovic and A. Zisserman, “All About VLAD,” in2013 IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 1578–1585

    2013

  32. [32]

    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). IEEE, 2016, pp. 770–778

    2016

  33. [33]

    Towards writer retrieval for historical datasets,

    M. Peer, F. Kleber, and R. Sablatnig, “Towards writer retrieval for historical datasets,” inDocument Analysis and Recognition - ICDAR 2023, 2023, pp. 411–427

    2023

  34. [34]

    Netvlad: CNN architecture for weakly supervised place recognition,

    R. Arandjelovic, P. Gronat, A. Torii, T. Pajdla, and J. Sivic, “Netvlad: CNN architecture for weakly supervised place recognition,” in2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, 2016, pp. 5297–5307

    2016