sonovice / smude Goto Github PK

Sauvola binarization eats dots sometimes. Thresholding for small object removal should be lowered.

I get following error after installing it with conda:

$ smude -o dewarped.png photo.png
Traceback (most recent call last):
  File "/usr/local/bin/smude", line 33, in <module>
    sys.exit(load_entry_point('smude==0.1.0', 'console_scripts', 'smude')())
  File "/usr/local/bin/smude", line 25, in importlib_load_entry_point
    return next(matches).load()
  File "/usr/local/Cellar/[email protected]/3.9.2_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/importlib/metadata.py", line 77, in load
    module = import_module(match.group('module'))
  File "/usr/local/Cellar/[email protected]/3.9.2_1/Frameworks/Python.framework/Versions/3.9/lib/python3.9/importlib/__init__.py", line 127, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
  File "<frozen importlib._bootstrap>", line 1030, in _gcd_import
  File "<frozen importlib._bootstrap>", line 1007, in _find_and_load
  File "<frozen importlib._bootstrap>", line 986, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 680, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 790, in exec_module
  File "<frozen importlib._bootstrap>", line 228, in _call_with_frames_removed
  File "/usr/local/lib/python3.9/site-packages/smude-0.1.0-py3.9.egg/smude/__init__.py", line 10, in <module>
    import torch
  File "/usr/local/lib/python3.9/site-packages/torch-1.8.0-py3.9-macosx-10.15-x86_64.egg/torch/__init__.py", line 196, in <module>
    from torch._C import *
RuntimeError: module compiled against API version 0xe but this version of numpy is 0xd

It would be useful, if the package can be installed with a setup.py that already installs all requirements. The first time I had problems to figure out the correct packages.

Additionally, I would like to use your tool direct on the command line and provide the images via parameters.

Something like:
smude -i input_image -o output_image

As far as I know, you need a main() function as entry point.

What do you think?

If I find some free time, I can create a PR.

Maybe this would be fixed if/when this thing goes on pip, but for now, doing the setup.py installation doesn't give me the machine-learning-related libraries and I have to get them manually.

Can I request a requirement.txt in the repo? (I know there's an environment.yml but pip install -r doesn't take that)

Having the ability to scale the output to have a fix-sized rastrum ("interline normalization") should be a rather low hanging fruit.

Q: Why would that be useful?
A: Most modern OMR pipelines use some sort of deep learning foo for musical object detection. These models usually utilize convolutional layers that are highly resolution dependent.

I already performed some tests and it works good in most cases. It's a great tool.

Is it be possible to deactivate the binarization step or recover the original images after dewarping? I would prefer to do the binarization manually afterwards to get a better result.

Right now the vanishing point computation (and some other things) depend on staff lines starting and ending with a bar line. Especially single instrument staves usually do not have such a starting line, resulting in various errors.

Basic idea:

1. Do staff line detection as already implemented
2. Draw blobs at the start and end of detected staff lines
3. Use Hough Transform on these "blob images" do detect the "non-existent bar lines"
4. Profit.

The "Region of Interest" mask extraction fails in a lot of cases. Should be overhauled or replaced altogether. Community ideas are welcome. ;)