TextBite is a human-annotated historical Czech document dataset for logical document segmentation and document layout analysis containing 8449 pages from Czech libraries.
| Mar. 14th, 2025 | The TextBite dataset has been published. |
The TextBite dataset consists of scanned historical Czech documents from various sources with diverse layouts. It includes simpler layouts, such as book pages and dictionaries, as well as more complex multi-column formats from newspapers, periodicals, and other printed materials. Additionally, part of the dataset contains handwritten documents, primarily records from schools and public organizations, introducing extra segmentation challenges due to their more loosely structured layouts.
In total, the dataset contains 8,449 annotated pages, from which 7,346 pages are printed and 1,103 are handwritten. The pages contain a total of 78,863 segments. The test subset contains 964 pages, of which 185 are handwritten. The annotations are provided in an extended COCO format. Each segment is represented by a set of axis aligned bounding boxes, which are connected by directed relationships, representing reading order. To include these relationships in the COCO format, a new top-level key relations is added. Each relation entry specifies a source and a target bounding box.
In addition to the layout annotations, we provide a textual representation of the pages produced by Optical Character Recognition (OCR) tool PERO-OCR. These come in the form of XML files in the PAGE-XML format, which includes an enclosing polygon for each individual textline along with the transcriptions and their confidences. Lastly, we provide the OCR results in the ALTO format, which includes polygons for individual words in the page image.
The dataset is publicly available at Zenodo.
| Dataset | Size | URL |
|---|---|---|
| TextBite Dataset | 11.7G | Download |
| Test Labels | 218.3M | Download |
| Baseline Models | 77.3M | Download |
TextBite provides four types of data assets:
- JPG images of all pages in their original resolution
- Bounding-box and relation annotations in COCO format for each JPG image
- OCR transcriptions with textline polygons in the PAGEXML format
- OCR transcriptions with word polygons in the ALTO format
The dataset is organized in the following directory structure:
├── coco
│ ├── test.json
│ ├── dev.json
├── images
│ ├── sample1.jpg
│ ├── ...
├── pagexml
│ ├── sample1.xml
│ ├── ...
├── alto
│ ├── sample1.xml
│ ├── ...
We propose evaluating logical page segmentation as a clustering problem, focusing on pixel-level segmentation instead of traditional text-based methods. Our approach excludes background pixels, considering only letter pixels, making evaluation OCR-independent and robust across segmentation techniques. Segmentation quality is measured using the Rand Index, comparing clustered text regions while ignoring background noise.
Evaluation can be performed using the /evaluation/eval_labeling.py script with the following parameters:
--ref-dir: Directory containing ground truth labels in .npy format.--hyp-dir: Directory with predictions in .json format. Each prediction file should contain a list of segments, where each segment is represented as a list of polygons. Example .json files can be seen in/data/json.
TextBite: A Historical Czech Document Dataset for Logical Page Segmentation
- Martin Kostelník ([email protected])
- Karel Beneš ([email protected])
- Michal Hradiš ([email protected])
ArXiv link: Arxiv
Citation:
@article{kostelnik2025textbite,
title={TextBite: A Historical Czech Document Dataset for Logical Page Segmentation},
author={Kosteln{\'\i}k, Martin and Bene{\v{s}}, Karel and Hradi{\v{s}}, Michal},
journal={arXiv preprint arXiv:2503.16664},
year={2025}
}