magistermilitum/Tridis_layout_manuscripts

A Unified Dataset for Codicological Document Layout Analysis

Dataset Description

This repository contains a large-scale, unified dataset for Document Layout Analysis (DLA) in historical manuscripts. It was created by harmonizing three distinct public corpora—e-NDP, CATMuS, and HORAE—which cover a wide range of document types from the 12th to the 17th century (administrative registers, literary manuscripts, printed books, and Books of Hours).

The key feature of this dataset is its hierarchical, multi-label annotation scheme. Unlike traditional flat object detection datasets, each object is annotated with its full path in a codicologically-informed ontology, enabling the training of sophisticated, semantically-aware models. For example, a simple initial is tagged with three labels: Initial, Initial_Manuscript, and Initial_Ms_Simple.

This dataset was created to train and evaluate the YOLO-gen foundational model and is intended to serve as a challenging new benchmark for the DLA community.

---

Training Data

The corpus harmonization was achieved through a custom hierarchical ontology described in the accompanying paper. The original annotations were transformed to Oriented Bounding Box (OBB) annotations in order to better detect complex manuscripts forms.

• e-NDP: A corpus of Parisian medieval registers (1326-1504) with a relatively homogeneous administrative layout.
  • Link: https://doi.org/10.5281/zenodo.7575693
• CATMuS: A diverse multi-class dataset derived from various medieval and modern sources (ca. 12th-17th c.), including administrative, literary, and printed documents.
  • Link: https://huggingface.co/datasets/CATMuS/medieval-segmentation
• HORAE: A corpus of richly decorated Books of Hours (ca. 13th-16th c.) with complex and artistic layouts.
  • Link: https://github.com/oriflamms/HORAE

Dataset Structure

The dataset is organized into three splits: train, validation (a 5% sample of the original training data), and test. Each split contains images and labels subfolders. The annotations follow the YOLO OBB (Oriented Bounding Box) format.

Label Format

Each .txt label file can contain multiple lines for a single object, corresponding to its hierarchical labels. The format for each OBB annotation line is:

<class_id> <x1> <y1> <x2> <y2> <x3> <y3> <x4> <y4>

Where (x_i, y_i) are the normalized coordinates of the four corners of the oriented bounding box.

Class Ontology

The V7 class structure was designed to harmonize the three disparate source datasets by reclassifying their original, non-interoperable labels into a single, coherent, and hierarchical ontology. The methodology is informed by established codicological vocabulary, allowing the model to learn relationships between layout elements.

The hierarchy is structured as follows:

• Parent Classes: These are the highest-level categories, representing fundamental codicological concepts like Text, Decoration, and Initial. They provide a shared, abstract vocabulary that is common across all document types.
• Intermediate Parents: In some cases, a second hierarchical level like Paratext or Initial_Manuscript was introduced. This serves to group elements by their function (e.g., separating main text from paratextual elements) or by their visual style (e.g., separating manuscript initials from printed ones), providing a necessary layer of nuance.
• Subclasses: These are the "leaf nodes" of the hierarchy (e.g., Text_Main, Paratext_Marginal, Initial_Ms_Simple). They are the direct mapping of the granular labels from the source corpora, preserving their original specificity within the new, structured framework.

This results in a multi-label dataset. Each object annotation includes its full hierarchical path. For instance, a marginal note from the e-NDP corpus is tagged with three labels: Text, Paratext, and Paratext_Marginal. This rich annotation scheme enables the training of powerful generalist models like YOLO-gen, which can understand the context and relationships between layout elements, and serves as a blueprint for future data harmonization efforts in the Digital Humanities.

ID	Class Name	Hierarchy	Source Corpora
0	Text	PARENT	-
1	Decoration	PARENT	-
2	Initial	PARENT	-
3	Marks	PARENT	-
4	Damage	PARENT	-
5	Numbering	PARENT	-
6	Paratext	Text > INTERMEDIATE	-
7	Initial_Manuscript	Initial > INTERMEDIATE	-
8	Initial_Printed	Initial > INTERMEDIATE	-
9	Text_Main	Text	E, C
10	Deco_Border	Decoration	H
11	Deco_Miniature	Decoration	H
12	Deco_Graphic	Decoration	C
13	Deco_LineFiller	Decoration	H
14	Initial_Ms_Simple	Initial > Initial_Manuscript	H
15	Initial_Ms_Decorated	Initial > Initial_Manuscript	H
16	Initial_Ms_Historiated	Initial > Initial_Manuscript	H
17	Initial_P_DropCapital	Initial > Initial_Printed	C
18	Marks_Quire	Marks	C
19	Marks_Stamp	Marks	C
20	Marks_Seal	Marks	C
21	Numbering_Page	Numbering	E, C
22	Damage_Generic	Damage	C
23	Damage_ScanArtifact	Damage	C
24	Paratext_Marginal	Text > Paratext	E, C, H
25	Paratext_Header	Text > Paratext	C
26	Paratext_List	Text > Paratext	E
27	Paratext_DateLine	Text > Paratext	E

• Source Corpora Key: E (e-NDP), C (CATMuS), H (HORAE).