cneud / ocr-gt Goto Github PK

View Code? Open in Web Editor NEW

70.0 7.0 11.0 130 KB

OCR & Ground Truth Resources

Home Page: https://cneud.github.io/ocr-gt/

optical-character-recognition ground-truth

ocr-gt's People

Contributors

Stargazers

Watchers

Forkers

jbaiter kba hulk89 wollmers stweil tboenig lena-hinrichsen aspnetcs farnazzeidi m3ssman

ocr-gt's Issues

IIT CDIP 1.0 (Illinois Institute of Technology Complex Document Information Processing Test Collection, version 1.0)

https://ir.nist.gov/cdip/README.txt

Generate web page for the YAML data?

This could be very simple, just a Jekyll page iterating through the entries in the YAML data and presenting them in a minimal HTML version at https://cneud.github.io/ocr-gt. It would even automatically rebuild for every change to the data. Or do you have another plan?

Make this a JSON file?

Since the majority of the information contained here has now been merged into kba/awesome-ocr and also due to the fact that Markdown tables are neither fun to edit nor specifically practical to reuse, I wonder: some information from here cannot easily be carried over to the awesome list without messing it up, but may still be relevant for some - would a JSON file make more sense perhaps? Or are there any other ideas, opinions on this? @kba @jbaiter

Cursive Greek

https://github.com/pharos-alexandria/ocr-greek_cursive/tree/master/gt

https://github.com/katabase/OCRcat

UB-Mannheim/Weisthuemer

https://github.com/UB-Mannheim/Weisthuemer

IMPACT github repo

@cneud,

Would you like to add links to these repos?

I notices that you are the one who uploaded the data to these repos a few years ago :-)

I assume that the first two are the same as the existing pol & spa data in your table.
Still, you can add them as alternative links.

Better metric: # of lines?

First of all, great work, thank you for putting together this list :-)

However, currently the only metric available in the comparison is the number of pages, which can be misleading. Archiscribe is listed with >3000 pages and e.g. the various IMPACT datasets with merely 50-200 pages. This makes it seem like the Archiscribe corpus is far bigger than it is: For each page, only one line was transcribed, while for IMPACT this was done for every line.
I think that, since for most OCR engines the line is the basis for training models, the comparison should reflect that.

https://groups.google.com/g/tesseract-dev/c/PBPHsyOm3SE

README: https://files.ifi.uzh.ch/cl/OCR19thSAC/readme.txt

License: CC-BY (https://files.ifi.uzh.ch/cl/OCR19thSAC/license.html)

The IAM Handwriting Database contains forms of handwritten English text which can be used to train and test handwritten text recognizers and to perform writer identification and verification experiments.

https://fki.tic.heia-fr.ch/databases/iam-handwriting-database