# CATMuS Medieval Segmentation: A SegmOnto-Based Layout Dataset for Medieval Manuscripts > CATMuS Medieval Segmentation provides 1,680 annotated manuscript pages spanning the 8th-16th centuries with 19 SegmOnto-based zone and line classes, offering a large-scale interoperable medieval layout resource. - Authors: Bevaya AI Team - Published: 2026-03-15 - Dataset: CATMuS Medieval Segmentation (HuggingFace) - Dataset: CATMuS Medieval HTR (HuggingFace) - Dataset: CATMuS Medieval (Zenodo) - Paper: Published (ICDAR 2024) - Paper: Preprint (HAL) - Source: Bevaya Labs, --- ## TL;DR CATMuS Medieval Segmentation is a 1,680-page layout dataset covering 159 medieval manuscripts (8th-16th century) in 10 languages, annotated with 19 classes following the SegmOnto vocabulary (13 zone types, 6 line types). It is released as a companion to the CATMuS Medieval HTR dataset, which was the primary focus of the ICDAR 2024 paper. Since the segmentation dataset has no standalone publication, this note draws on the ICDAR 2024 paper for context on dataset construction, sources, and methodology. The segmentation dataset provides hierarchical block/line annotations with polygons, bounding boxes, and baselines. ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$ The primary contribution is a large-scale, multi-project, interoperable dataset for medieval manuscript analysis. The segmentation dataset is a companion release providing layout annotations derived from the HTR ground truth creation pipeline. **Secondary:** $\Psi\_{\text{Evaluation}}$ The paper devotes a full section to benchmark split design (general vs. feature-based) and HTR baseline experiments with two engines across three seeds, reporting CER with standard deviations. ## What is the motivation? Digitization initiatives have made thousands of medieval manuscripts available as images, but most lack machine-readable text. Handwritten Text Recognition (HTR) is the key technology for conversion, yet building consistent ground truth across projects is difficult due to conflicting philological traditions, varying transcription standards, and project-specific annotation practices. Prior to CATMuS, no single dataset provided cross-century, cross-language, and cross-script evaluation capability for medieval Latin-script manuscripts. Each project worked in isolation with incompatible conventions. For layout analysis specifically, medieval manuscripts present challenges that modern document datasets do not cover: marginal glosses, drop capitals, music notation, quire marks, seals, and other elements unique to historical codices. The SegmOnto vocabulary was designed to provide a shared ontology for these elements across projects. ## What is the novelty? CATMuS Medieval is a **macro-dataset** assembled from 14 contributing repositories across multiple institutions. The core contribution is the unification effort: harmonizing transcription guidelines, segmentation practices, and metadata across projects that previously used incompatible standards. The dataset construction followed three phases: 1. **Phase 1:** CREMMA and HTRomance projects formulated primary transcription guidelines for literary manuscripts in book scripts (Old French, Latin, Castilian, Italian). 2. **Phase 2:** The DEEDS project joined, adding Latin cartulary material. 3. **Phase 3:** Adoption by the Carthusian Monastery of Herne team and Middle English researchers. Segmentation was performed primarily in eScriptorium using the BLLA model, with corrections. Transkribus data was repolygonized and post-corrected for compatibility. The segmentation dataset applies SegmOnto vocabulary for both zone-level and line-level classification. ### Transcription Guidelines A significant part of the paper's contribution is the **graphemic transcription approach**, which enables interoperability across projects: - **Allograph normalization:** Many-to-one mapping (e.g., round s and long f merged; round r and short r merged). - **Abbreviations preserved:** Treated as an NLP task, not an HTR task. Diacritics in Unicode decomposed form (NFD). - **u/v and i/j normalization:** The "Ramist" distinction is disregarded; v is transcribed as u, j as i. - **Punctuation standardized:** Reduced to four signs: `.` (full stop), `;` (double signs), `,` (comma), `?` (question mark). - **Word segmentation:** Semantic (lexical spaces) rather than imitative, addressing inconsistencies across medieval scribal practices. These choices allow manuscripts from different projects, languages, and centuries to coexist in a single training set without conflicting conventions. ### SegmOnto Classes (19 total) **Zone types (13):** | Zone Class | Description | | :--- | :--- | | MainZone | Primary text area | | MarginTextZone | Text in margins (glosses, annotations) | | DropCapitalZone | Decorated initial letters | | GraphicZone | Illustrations or decorative elements | | MusicZone | Musical notation | | NumberingZone | Folio/page numbering | | QuireMarksZone | Quire signatures or catchwords | | RunningTitleZone | Running headers/titles | | SealZone | Wax or parchment seals | | StampZone | Library or ownership stamps | | TitlePageZone | Title pages (primarily incunabula) | | DamageZone | Areas of physical damage | | DigitizationArtefactZone | Scanning artifacts (rulers, color bars) | **Line types (6):** | Line Class | Description | | :--- | :--- | | DefaultLine | Standard text lines | | HeadingLine | Section headings or rubrics | | DropCapitalLine | Lines containing drop capitals | | InterlinearLine | Interlinear glosses or corrections | | MusicLine | Lines of musical notation | | TironianSignLine | Lines containing Tironian shorthand | Note: The class inventory above is sourced from the [HuggingFace dataset card](https://huggingface.co/datasets/CATMuS/medieval-segmentation). The ICDAR 2024 paper mentions region and line classes as metadata but does not enumerate the full SegmOnto vocabulary. ## What experiments were performed? The ICDAR 2024 paper evaluates HTR baselines (not layout analysis) on the companion CATMuS Medieval HTR dataset. Two HTR engines were benchmarked across two split strategies: - **General split** (90/5/5): Every manuscript appears in all partitions. This tests in-domain generalization. - **Feature-based split** (~85/7/7): Entire manuscripts are assigned to a single partition. This tests out-of-domain generalization to unseen manuscripts, scripts, and centuries. | Engine | Split | Train Time (min) | Val CER (%) | Test CER (%) | Space CER (%) | | :--- | :--- | :--- | :--- | :--- | :--- | | Kraken v4.3.10 | General | $2112 \pm 163$ | $5.7 \pm 0.07$ | $4.7 \pm 0.06$ | $1.0 \pm 0.02$ | | Kraken v4.3.10 | Feature | $1464 \pm 238$ | $6.8 \pm 0.16$ | $13.1 \pm 0.24$ | $2.7 \pm 0.06$ | | Pylaia v1.1.0 | General | $308 \pm 47$ | $9.1 \pm 0.63$ | $8.4 \pm 0.73$ | $1.8 \pm 0.11$ | | Pylaia v1.1.0 | Feature | $295 \pm 78$ | $11.3 \pm 0.24$ | $21.2 \pm 0.92$ | $3.8 \pm 0.06$ | The large CER gap between general and feature-based splits (e.g., 4.7% vs. 13.1% for Kraken) highlights the difficulty of generalizing to unseen manuscripts. Space-related errors are a notable component, reflecting inconsistent word separation practices across medieval languages and centuries. No layout analysis models were evaluated on the segmentation dataset. The segmentation annotations serve as ground truth for line extraction in the HTR pipeline but have not been benchmarked independently for object detection or instance segmentation tasks. ## What are the outcomes/conclusions? The CATMuS Medieval project demonstrates that cross-institutional collaboration on annotation standards is feasible for historical manuscripts, producing a dataset of 208 documents, 165,347 lines, and 5.77 million characters across 10 languages and 9 centuries. Key limitations: - **Imbalance:** The Carthusian Monastery of Herne project alone contributes 27% of lines, creating overrepresentation of 14th-century Dutch Textualis. Old French, Latin, and Castilian together exceed 50% of the data. - **Rare languages:** Venetian (2 documents), Navarese (1), and Old English (1) have minimal coverage. - **Image quality variation:** 30 of 203 documents are grayscale (from microfilm), with varying resolution across sources. - **Layout coverage:** The segmentation dataset (1,680 pages, per the HuggingFace release) is smaller than the full HTR dataset (165K+ lines across 208 documents), suggesting not all pages have full region-level annotations. - **No layout baselines:** The paper does not evaluate layout detection performance, so it is unknown how well modern detectors handle the SegmOnto class set on this data. - **No segmentation pipeline code:** The XML-to-Parquet conversion, SegmOnto remapping, and polygon extraction steps used to produce the segmentation dataset are not released as code. ## Reproducibility ### Models Not applicable. This is a dataset paper. No layout models are proposed or released. The HTR baselines use default Kraken and recommended Pylaia architectures without modification. ### Algorithms HTR training hyperparameters reported in the paper: - **Batch size:** 64 - **Seeds:** 21, 42, 84 (three runs per configuration) - **Early stopping:** 5-epoch patience - **Learning rate:** $1 \times 10^{-4}$ (Kraken), $5 \times 10^{-4}$ (Pylaia) - **Deterministic mode:** Enabled - **Configuration:** Default specifications for Kraken; recommended specifications for Pylaia (per Maarand et al.) No layout-specific training algorithms are reported. ### Data The segmentation dataset is available on HuggingFace in Parquet format (statistics from the [HuggingFace dataset card](https://huggingface.co/datasets/CATMuS/medieval-segmentation)): - **Train:** 1,340 pages - **Validation:** 191 pages - **Test:** 158 pages - **Total:** 1,680 pages from 159 manuscripts Each record includes the page image, bounding boxes, polygon masks, baselines, SegmOnto category labels, block/line type, parent-child hierarchy, manuscript shelfmark, century, and source project. The broader HTR dataset is available separately on HuggingFace (`CATMuS/medieval`) and Zenodo, containing 165,347 extracted line images with transcriptions and rich metadata (language, script type, genre, century). **Annotation:** - **Tool:** eScriptorium (primary), Transkribus (Middle Dutch subset) - **Annotators:** 45+ contributors across 14 source projects - **Process:** Automatic segmentation via BLLA model, followed by manual correction - **Format:** Polygonal masks with baselines; hierarchical (lines linked to parent blocks) ### Evaluation No layout evaluation metrics or baselines are reported for the segmentation dataset. HTR evaluation uses Character Error Rate (CER): $$ \text{CER} = \frac{S + D + I}{N} $$ where $S$, $D$, and $I$ are the number of substitutions, deletions, and insertions respectively, and $N$ is the number of characters in the reference. The paper also reports a "space-related" CER isolating spacing errors, which is relevant given medieval word separation variability. ### Hardware HTR training was performed on NVIDIA RTX 8000 GPUs with 12 CPU cores per run. Kraken training took substantially longer than Pylaia ($\sim$2100 min vs. $\sim$300 min on the general split). No layout-specific training was reported. ## BibTeX ``` ```bibtex @inproceedings{clerice2024catmus, author = {Cl{\'e}rice, Thibault and Pinche, Ariane and Vlachou-Efstathiou, Malamatenia and Chagu{\'e}, Alix and Camps, Jean-Baptiste and Gille Levenson, Matthias and Brisville-Fertin, Olivier and Boschetti, Federico and Fischer, Franz and Gervers, Michael and Boutreux, Agn{\`e}s and Manton, Avery and Gabay, Simon and O'Connor, Patricia and Haverals, Wouter and Kestemont, Mike and Vandyck, Caroline and Kiessling, Benjamin}, title = {{CATMuS Medieval: A multilingual large-scale cross-century dataset in Latin script for handwritten text recognition and beyond}}, booktitle = {Document Analysis and Recognition -- ICDAR 2024}, pages = {174--194}, year = {2024}, publisher = {Springer}, address = {Athens, Greece}, doi = {10.1007/978-3-031-70543-4_11} } ``` ```