# TexBiG: A Dataset for Analysing Complex Document Layouts in the Digital Humanities > TexBiG provides 52K+ expert polygon annotations across 19 classes for historical documents from the late 19th and early 20th centuries, with inter-annotator agreement measured by Krippendorff's Alpha. - Authors: Bevaya AI Team - Published: 2026-03-15 - Paper: Published (DAGM GCPR 2022) - Dataset: TexBiG Dataset (Zenodo) - Code: VSR-TexBiG Code (GitHub) - Project: Annotation Guidelines (PDF) - Model: Pretrained Weights (Zenodo) - Source: Bevaya Labs, --- ## TL;DR TexBiG (Text-Bild-Gefuge, "Text-Image-Structure") is a document layout analysis dataset for historical documents from the late 19th and early 20th centuries. It provides over 52,000 expert polygon annotations across 19 classes, with inter-annotator agreement measured using Krippendorff's Alpha. Each document image was annotated by at least two independent annotators. Baselines are provided using Mask R-CNN and VSR. --- ## What kind of paper is this? **Dominant:** $\Psi\_{\text{Resource}}$ The headline contribution is the dataset itself: its annotation guidelines, 19-class taxonomy, polygon annotations, and inter-annotator agreement evaluation. The paper is organized around data curation and quality assessment. **Secondary:** $\Psi\_{\text{Evaluation}}$ The paper benchmarks Mask R-CNN and VSR on the dataset and uses Krippendorff's Alpha to formally quantify annotation reliability, which is unusual for layout datasets. ## What is the motivation? Digital humanities researchers work with historical documents that differ substantially from the born-digital, English-language STEM papers that dominate existing DLA datasets. Late 19th and early 20th century publications exhibit complex text-image arrangements, varied typography, decorative elements, and advertisements interleaved with body content. Existing datasets either lack the class granularity needed for these documents or do not provide polygon-level annotations that can capture the irregular region shapes common in historical layouts. The authors also note that most DLA datasets lack formal inter-annotator agreement measurements, making it difficult to assess annotation reliability. TexBiG addresses both gaps: fine-grained polygon annotations for historical documents with formal agreement metrics. ## What is the novelty? 1. **19-class polygon annotations.** Instance segmentation annotations (bounding boxes + polygons/masks) in COCO format across 19 classes covering both textual and non-textual document elements. 2. **Expert annotation with formal agreement.** Each document image was annotated by at least two independent expert annotators. Inter-annotator agreement is measured using Krippendorff's Alpha ($\alpha$), providing a principled estimate of annotation reliability. Krippendorff's Alpha is defined as: $$\alpha = 1 - \frac{D\_o}{D\_e}$$ where $D\_o$ is the observed disagreement between annotators and $D\_e$ is the disagreement expected by chance. Values of $\alpha \geq 0.8$ are generally considered reliable, while $0.667 \leq \alpha < 0.8$ allows tentative conclusions. 3. **Historical document focus.** The dataset draws from 7 publicly available German-language publications from the late 19th and early 20th centuries, covering a range of layout complexity from simple book pages to magazine-style layouts with embedded images, advertisements, and decorative elements. 4. **OCR data included.** Tesseract-generated OCR output is provided alongside the annotations to support multimodal models like VSR, though the authors note this OCR is imperfect. ## What experiments were performed? Two baseline models were evaluated: - **Mask R-CNN** (via Detectron2): standard instance segmentation baseline - **VSR** (Vision-Semantics-Relations): multimodal model using both visual features and OCR text, pre-trained on PubLayNet Results are reported as mAP on the test split. The paper DOI (10.1007/978-3-031-16788-1_22) provides full benchmark numbers (the paper itself is not open access). ## What are the outcomes/conclusions? TexBiG fills a gap for historical document layout analysis with a well-annotated, multi-class dataset. The formal inter-annotator agreement evaluation via Krippendorff's Alpha sets it apart from most DLA datasets, which either report no agreement metrics or use ad-hoc measures. The dataset has been adopted by subsequent work (e.g., AnnoPage re-annotates TexBiG pages under its own schema). ### Limitations - The paper is behind a paywall (Springer DAGM GCPR), limiting accessibility. - The GitHub repository for the VSR benchmark code has no license file. - The dataset is relatively small by deep learning standards (7 books), though the annotation density (52K+ instances) is high. - OCR data is Tesseract-generated and imperfect. ## Mapping to Matter vs. Meaning Framework How does TexBiG's 19-class taxonomy map to our [Matter vs. Meaning framework](../matter-vs-meaning/)? | TexBiG Class | Visual Primitive | Logical Role | Notes | | :--- | :--- | :--- | :--- | | **paragraph** | `Text` | `Body` | Standard prose content. | | **heading** | `Text` | `SectionHeader` | Section headings. | | **sub-heading** | `Text` | `SectionHeader` | Subsection headings. | | **column title** | `Text` | `SectionHeader` / `Title` | Column-level titles in multi-column layouts. | | **author** | `Text` | `Author` | Author names. | | **caption** | `Text` | `Caption` | Descriptions of images/tables. | | **footnote** | `Text` | `Footnote` | Bottom-of-page notes. | | **header** | `Text` | `PageHeader` | Running headers. | | **footer** | `Text` | `PageFooter` | Running footers. | | **page number** | `Text` | `PageNumber` | Folio numbers. | | **editorial note** | `Text` | `Annotation` | Editor's notes, commentary. | | **equation** | `Formula` | `DisplayEquation` | Mathematical content. | | **table** | `Table` | *(primitive only)* | Tabular data. | | **image** | `Image` | `Figure` | Photographs, illustrations. | | **logo** | `Image` | `Logo` | Publisher or organizational logos. | | **advertisement** | `Image` | `Advertisement` | Commercial advertisements. | | **decoration** | `Structure` | *(decorative)* | Ornamental elements, borders, vignettes. | | **frame** | `Structure` | *(structural)* | Page frames, column borders. | | **noise** | `Structure` | *(artifact)* | Scanner artifacts, stains, damage. | **Coverage strengths:** TexBiG provides a well-balanced taxonomy for historical documents. It distinguishes `Logo` from generic `Image`, separates `Advertisement` from editorial content, and includes `decoration`, `frame`, and `noise` classes that are important for historical material but absent from most modern-document datasets. The `editorial note` class is unique among DLA datasets. **Coverage gaps:** No explicit classes for `ListItem`, `BibEntry`, `Abstract`, `Blockquote`, `Code`, `OpticalCode`, or form primitives. No relation annotations (reading order, hierarchy). ## Reproducibility ### Models - **Mask R-CNN:** Standard Detectron2 implementation. Pretrained weights released on Zenodo. - **VSR:** Multimodal model from Hikvision's DAVAR-Lab-OCR, pre-trained on PubLayNet. Code adapted for TexBiG and released on GitHub. ### Algorithms - **Framework:** Detectron2 for Mask R-CNN; mmdetection 2.11.0 for VSR. - **Pre-training:** VSR initialized from PubLayNet-pretrained weights. - **Training details:** Optimizer, learning rate, batch size, epochs, and augmentation details are not available outside the paywalled paper. ### Data - **Source:** 7 publicly available German-language historical publications (late 19th to early 20th century). - **Total annotations:** 52,000+ instances across 19 classes. - **Format:** COCO JSON with polygon masks and bounding boxes. Train/val/test splits provided. - **Annotation process:** Manual expert annotation using a detailed guideline (provided as PDF on Zenodo). At least 2 annotators per document. Agreement measured with Krippendorff's Alpha. - **OCR:** Tesseract-generated text provided in DAVAR format for VSR compatibility. - **License:** CC-BY-4.0 for both annotations and images. Attribution required. - **Access:** Zenodo ([doi:10.5281/zenodo.6885144](https://zenodo.org/records/6885144)). ### Evaluation - **Metric:** mAP (COCO-style) for detection and segmentation. - **Agreement:** Krippendorff's Alpha for inter-annotator reliability. - **Baselines:** Mask R-CNN and VSR. - Full benchmark results in the published paper (Springer, paywalled). ### Hardware - NVIDIA Tesla V100-PCIE (16 GB), Ubuntu 18.04. ## BibTeX ```bibtex @inproceedings{tschirschwitz2022texbig, title={A Dataset for Analysing Complex Document Layouts in the Digital Humanities and its Evaluation with Krippendorff's Alpha}, author={Tschirschwitz, David and Klemstein, Franziska and Stein, Benno and Rodehorst, Volker}, booktitle={DAGM German Conference on Pattern Recognition}, pages={354--374}, year={2022}, publisher={Springer}, doi={10.1007/978-3-031-16788-1_22} } ```