[cv2] Completed type stubs for the following functions and their shared parameters:

[Avasam]

• cv2.norm
• cv2.resize
• cv2.cvtColor
• cv2.imread
• cv2.inRange
• cv2.imwrite
• cv2.calcHist
• cv2.normalize
• cv2.compareHist
• cv2.matchTemplate
• cv2.minMaxLoc
• cv2.bitwise_and
• ndarray

I filled in the type stubs based on our usage in https://github.com/Toufool/Auto-Split (Toufool/AutoSplit/pull/92) and what documentation I could find online.
Fixes microsoft/pylance-release/issues/2089

[ghost]

All CLA requirements met.

[Avasam]

Any update? This has been stalling for over 4 months without any acknowledgment.

[bschnurr]

Sorry forgot about this PR. I was hoping this OpenCV PR was going to solve everything.
https://github.com/opencv/opencv/pull/20370/files

Only suggestion would be to use the OpenCV alias names for parameters like Mat instead ndarray directly.
In that PR I linked they have a bit table of alias names.

I'm no openCV expert, not sure is Mat is exactly an ndarry though.

[Avasam]

Sorry forgot about this PR. I was hoping this OpenCV PR was going to solve everything. https://github.com/opencv/opencv/pull/20370/files

Oh looks like they might be adding some autogenerated type stubs soon. That would be nice.

Only suggestion would be to use the OpenCV alias names for parameters like Mat instead ndarray directly. In that PR I linked they have a bit table of alias names.

The type alias don't exist yet, but I can easily do Mat = np.ndarray if you'd like.
This PR also specified the generic types of ndarray (numpy.ndarray[int, numpy.dtype[typing.Any]]) so it doesn't result in ndarray[Unknown] (partially Unknown) errors with strict typing. If you're hesitant about that, I can split that in a different PR.

I'm no openCV expert, not sure is Mat is exactly an ndarry though.

Not an expert either, but I know that cv2 (opencv-python) uses numpy arrays directly, there's no "Mat" (Mat class definition in C++ : https://docs.opencv.org/4.5.5/d3/d63/classcv_1_1Mat.html)

[bschnurr]

does pandas dtype example help?
https://github.com/pandas-dev/pandas/blob/9b4f03e30a5b4d2c073d4f0c357fe1d02e7cb05d/pandas/_typing.py#L129

maybe numpy.ndarray[int, numpy.dtype[object]] is better following typeshed notes below.

https://github.com/python/typeshed/blob/master/CONTRIBUTING.md#conventions
"Note that Any is not the correct type to use if you want to indicate that some function can accept literally anything: in those cases use object instead."

[bschnurr]

More discussion here. python/typing#516

[bschnurr]

does numpy.DTypeLike help? I see pandas using that a lot

[Avasam]

Thanks for the references !

Numpy themselves use dtype[Any]

object is not compatible, but I guess it should be np.generic

does numpy.DTypeLike help? I see pandas using that a lot

So then: Mat = np.ndarray[int, np.dtype[np.generic]]

[Avasam]

Or even Mat = npt.NDArray[np.generic] which is equivalent to Mat = np.ndarray[Any, np.dtype[Any]]. This way we don't add Any directly to this code base, and don't restrict the Mat to int/float (in case that causes issues with co/contravariance). While still avoiding strict type checker issues due to Unknown or missing generic params.

[bschnurr]

Thanks for the contributions. Going forward I'd like to fix a bunch of the return types too.

[bschnurr]

hmm. np.generic might only work for python version 3.9.

[Avasam]

hmm. np.generic might only work for python version 3.9.

Is something not working? numpy.generic has been added 17 years ago. However the type stub has only been created around 3 years ago.

[bschnurr]

still need to test on python 3.8 and 3.7... just saw similar issue with panda's test failing

[neur1n]

Why the dts and normType arguments in some functions (such as normalize) changed from dst and norm_type? I think these new naming is different from the API's.