368 MB
3 files
Updated about 1 month ago
Name
Size
data
.gitattributes2.42 kB
xet
README.md4.18 kB
xet
README.md

ATR benchmark - Page/paragraph level

Dataset Description

Dataset Summary

The ATR benchmark dataset is a multilingual dataset that includes 463 document images, at page or paragraph level. This dataset has been designed to test ATR models and combines data from several public datasets:

Images are in their original size.

Split

dataset images language
BnL Historical Newspapers 3 German
CASIA-HWDB2 10 Chinese
Churro 290 Multi-lingual
DAI-CRETDHI 10 French
DIY History - Social Justice 20 English
Esposalles 10 Catalan
FINLAM - Historical Newspapers 10 English / French
Horae - Books of hours 10 Latin
IAM 10 English
NDLOCR 10 Japanese
NorHand v3 10 Norwegian
OpenITI 10 Arabic
Marius PELLET 10 French
QARI 10 Arabic
RASM 10 Arabic
READ-2016 10 German
RIMES 10 French
ScribbleLens 10 Dutch

Dataset Structure

Data Instances

{
  'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size={img size} at 0x1A800E8E190,
  'text': '{transcription}'
}

Data Fields

  • image: a PIL.Image.Image object containing the image. Note that when accessing the image column (using dataset[0]["image"]), the image file is automatically decoded. Decoding of a large number of image files might take a significant amount of time. Thus it is important to first query the sample index before the "image" column, i.e. dataset[0]["image"] should always be preferred over dataset["image"][0].
  • dataset: the name of the original dataset.
  • text: the label transcription of the image.
  • level: full document page or a single paragraph.
Total size
368 MB
Files
3
Last updated
Oct 22
Pre-warmed CDN
US EU US EU

Contributors