Accuracy and speed tests

[amitdo]

Related to discussion in: #707 (comment).

@stweil,

Can you contribute these images and their GT and release them under Apache 2.0 license?

Then, can you make a new daily CI test (for Ubuntu at least) that will include these images?

UNLV images could also be used as part of an accuracy test with CI.

[stweil]

All images and the ground truth are already available online under the free license Public Domain Mark 1.0.

The script for the evaluation is also available online (do5.sh here, currently uses local filenames), and so is the list of GT files.

The URLs for a GT filename follow this pattern:

41989179X_0036.txt is at https://digi.bib.uni-mannheim.de/fileadmin/digi/41989179X/gt/41989179X_0036.txt, the image is at https://digi.bib.uni-mannheim.de/fileadmin/digi/41989179X/max/41989179X_0036.jpg.

[amitdo]

According to this report, there was a huge regression in accuracy and speed in version 3.03 compared to 3.02.02.

We need to evaluate Tesseract weekly using the UNLV dataset, in order to make sure there are no regression in performance. The legacy engine and the CRNN engine should be tested separately. This is very important TODO.