Why __init__ is required?
I think Mat is RealisticInfoArray from the user guide except that Mat should wrap only numeric types. For Python objects and strings it should raise AssertionError

And it is better to explicitly add wrap_channels argument to __new__ with the documentation string.

It make sense to keep instance custom initialization in __init__ instead of __new__.

• https://docs.python.org/3/reference/datamodel.html
• debugging purposes
• code autocompletion works better with __init__

I thought about pre-checks too, but they can be added later. At least to keep this patch minimal. Currently such cases are catched in cv2.cpp anyway.

2 practical use cases to test:

1. Create a Mat wrapper for ndarray and try to use one of the NumPy ufuncs e.g.:

data = np.arange(10)
mat_data = cv.Mat(data)
np.testing.assert_equal(2 * data, 2 * mat_data)

2. Should 2 Mat objects constructed with different value of wrap_channels property be treated as equals?

data = np.ones((10, 10, 3))
mat_wrapped = cv.Mat(data, wrap_channels=True)
mat_simple = cv.Mat(data)
self.assertEqual(mat_wrapped, mat_simple) # ???

@VadimLevin Thank you for the review!

I added ufunc test case.
Comparison is tricky for now (numpy == returns an array which can't be casted to boolean result without .any/.all calls), so I would like to postpone introducing of custom comparison overloads.