Very good overall performance, but this one fails?

[jbarth-ubhd]

Here the original image: https://digi.ub.uni-heidelberg.de/diglitData/v/blaeu1655bd6_-_00_129.tif

here the image fed into sbb-textline (binarized etc): https://digi.ub.uni-heidelberg.de/diglitData/v/blaeu1655bd6_-_00_129-binarized.png

and here the detected segments:

model used:

-rw-r--r-- 1 jb users  458969872 Dec 10  2019 /usr/local/ocrd_models/sbb/textline/model_page_mixed_best.h5
-rw-rw-r-- 1 jb users 1194551551 Feb 13  2020 /usr/local/ocrd_models/sbb/textline/models.tar.gz
-rw-r--r-- 1 jb users  458970960 Jun 26  2019 /usr/local/ocrd_models/sbb/textline/model_strukturerkennung.h5

[cneud]

@jbarth-ubhd weird, I have not seen any segmentation results like this coming from the tool. Can you attach the PAGE-XML as well plz?

(cc @vahidrezanezhad)

[jbarth-ubhd]

I'll let in through a second time, just to be sure...

[jbarth-ubhd]

yes, second try, same, result. Complete ocr-d workflow results:

https://digi.ub.uni-heidelberg.de/diglitData/v/blaeu.tgz

[cneud]

Thanks for providing the test data. I can also confirm this via Aletheia. The issue seems to be with the region segmentation - where there are regions detected (see e.g. the marginalia on the left hand side), the textline segmentation actually works ok-ish. We will have a look what's wrong here!

Btw, just in case you missed this recent announcement in the OCR-D Chat:

[...] this is just an ahead notice that SBB/SPK will soon phase out the [development of the] sbb-textline-detector tool in favour of a new version of a layout detection tool which brings numerous improvements, such as e.g.

• more granular CLI/API access,
• additional layout elements being detected (marginalia, headlines, initials) and
• generally much improved performance.

We aim to provide the new tool with OCR-D compliant interfaces via our regular GitHub https://github.com/qurator-spk [asap]

[jbarth-ubhd]

I'm looking forward for this! Thanks for inspecting.

[cneud]

Dear @jbarth-ubhd, I found some time to investigate this further and with the current version of sbb-textline-detector, while I do get the same problem with the RGB image, when using the binarized image, the segmentation result actually looks fine (see attachments).

Using RGB image

Using binarized image (regions)

Using binarizd image (lines)

I will also try again with our new (but still work-in-progress) segmentation tool which prefers non-binarized images as input and post results here.