Refactor dewarping to be runnable

OCR-D / ocrd_anybaseocr

DFKI Layout Detection for OCR-D

Apache License 2.0

48 stars 12 forks source link

Refactor dewarping to be runnable #62

Closed kba closed 4 years ago

kba commented 4 years ago

Integrates pixp2pixhd as a submodule and tricks it into behaving like a subpackage
Refactored the dewarp processor:
- Fix broken option passing to pix2pixHD
- remove second output file group
- simplified ocrd-tool.json to only require the path to the trained pix2pixHD/pytorch model
Added a simple unit test that fails with the dfki-testdata that was specifically created for this project, because of course it fails

So I still cannot say whether the pix2pixHD inference does anything useful, I'll try to fix the unit test with appropriate test data asap.

stweil commented 4 years ago

ocrd_anybaseocr is currently broken with the OCR-D reference Ubuntu 18.04 because it installs a Keras package which is too new. This is fixed here. I'd appreciate if such small fixes could be applied without too much delay. The fix for the Keras dependency is even trivial enough to apply it without waiting for a pull request.

kba commented 4 years ago

This is fixed here. I'd appreciate if such small fixes could be applied without too much delay. The fix for the Keras dependency is even trivial enough to apply it without waiting for a pull request.

Well, we all would appreciate that but the keras fix was not the issue here. I didn't know that this was such a pressing issue for you - why didn't you open an issue or send a PR? I'll fix it in a separate PR, can't test this PR now.

stweil commented 4 years ago

I only noticed the Keras issue with ocrd_anybaseocr today, fixed it locally and was preparing a pull request when I noticed that it is already part of the one here. So it is not a pressing issue for me, but of course it was nagging for anyone trying OCR-D. Thank you for your fix!