kraken 5.2.4 on eScriptorium recognition artefacts

[johnlockejrr]

I'm not sure if is eScriptorium or kraken related, I just want to poin out, same model, same image on different installs:

1. eScriptorium with kraken 4.3.13 (python 3.8)

2. eScriptorium with kraken 5.2.4 (python 3.10)

Both segmentation and recognition models were trained on kraken 5.2.4

[dstoekl]

looks like shapely

[dstoekl]

the polygon is too big and the recognizer wasn't trained on lines where the letters are only a quarter of the line height.

[johnlockejrr]

Here is on kraken 4.x, same model.

[johnlockejrr]

The models I used in this test:
mcdonald.zip

[mittagessen]

It isn't a model issue but the polygonization is wrong. I'll have a look. The rotation code changed between 4.x and 5.x so it's either that or other shapely shenanigans.

[mittagessen]

Could you also send me the image file and any ALTO/PageXML you've got? It's difficult to debug without being able to run a test case.

[johnlockejrr]

export_doc23_memar_marqah_mcdonald_alto_202405131147.zip
Sure, here it is the image with ALTO (from 4.x)

[mittagessen]

Thanks. It's mostly so I can make sure the baselines are identical.

[johnlockejrr]

Any update on this matter?

[mittagessen]

Apparently, the error persists on some other image data.

[mittagessen]

Nope, not true after all. Just crappy output of the polygonizer.