add qurator-spk/sbb_binarization

[kba]

Pixelwise binarization with selectional auto-encoders in Keras

@jbarth-ubhd has an example of the result in OCR-D/ocrd-website#172 (comment)

[stweil]

Looks good. Are there already models available for production? Then a rule install-models-sbb-binarize to download one or several models would be nice.

[bertsky]

I second @stweil's comment about packaging models. Or can we do this via package_data and the upcoming resource discovery feat?

[kba]

Pre-trained models are available from https://qurator-data.de/sbb_binarization/

I'll add an install-models rule for this as well.

As for bundling with the tool, I'll have to finish that in core first but then yes, certainly, they could be placed in one of the defined locations and model parameter could be a relative path. For now it must be absolute.

[bertsky]

As for bundling with the tool, I'll have to finish that in core first but then yes, certainly, they could be placed in one of the defined locations

But the Python package on PyPI should not be polluted with such large models. We could use setuptools' extra_requires feature to build packages including them from our makefile.

and model parameter could be a relative path. For now it must be absolute.

Couldn't we add an environment variable (say SBB_BINARIZATION_MODELS) in the meantime? Absolute paths make sharing makefile configurations across hosts impossible.

[kba]

Then a rule install-models-sbb-binarize to download one or several models would be nice.

32285fa

Couldn't we add an environment variable (say SBB_BINARIZATION_MODELS) in the meantime? Absolute paths make sharing makefile configurations across hosts impossible.

Yes but then I'd suggest to implement it in a processor-independent way, see OCR-D/spec#176

[bertsky]

Yes but then I'd suggest to implement it in a processor-independent way, see OCR-D/spec#176

That's not the same thing, though. Yours is an environment variable for a specific parameter (which is entirely new AFAIK). My suggestion was to follow the example of TESSDATA_PREFIX, OCROPUS_DATA, CORASVANN_DATA to allow passing a reference point in the search paths for relative filenames. (Not standardized either, of course – but usable already.)

[kba]

(Not standardized either, of course – but usable already.)

I realize this is more specific but it would not be much harder to implement in core for all the processors to make use of.

I can implement (=steal from cor-asv-ann ;-)) SBB_BINARIZATION_DATA in the short term, I'll send a PR.

[bertsky]

I can implement (=steal from cor-asv-ann ;-)) SBB_BINARIZATION_DATA in the short term, I'll send a PR.

That would be great. In particular for derived Docker images :-)

[kba]

I can implement (=steal from cor-asv-ann ;-)) SBB_BINARIZATION_DATA in the short term, I'll send a PR.

qurator-spk/sbb_binarization#6