LADaS, created by the ALMANaCH team-project at Inria, continued in partnership with other researchers, is a multidocuments diachronic layout analysis dataset. This dataset includes:
- Monographs from the Bibliothèque Nationale de France (17th century - today);
- PhD Thesis, in various fields (not only STEM, 20th-21st century);
- Selling Catalogs (for manuscripts and art pieces), in various fields (18th-20th century);
- Noisy digitization (with fingers for example, 20th-21st century);
- Academic papers (mostly Humanities and Social Sciences) (19th-21st century);
- Magazines about technologies and video games, from 1920s to 2010;
- Misc stuff found here and there.
The data are in YoloV8 txt format (class center_x center_y width height).
The script in document is mostly Latin script, and language is mostly French with some representation of the main western academic languages.
Label Annotation have been conducted using the SegmOnto vocabulary. An annotation guide is available here.
Last update of the plot: 15/12/2023
The data can be found in ./data. Each subset is present in its own subset folder if you want to train cross-genre.
A script, collate.sh allows for having a single directory with train/dev/test folders for YoloV8 training.
See the CITATION.CFF file
Thibault Clérice ( th[a-z]+.cle[a-z]+ [at] inria.fr)