Type annotation issues with opencv-python 4.9.0.80 about opencv HOT 4 CLOSED

• The borderValue argument to cv2.erode is annotated as cv2.typing.Scalar, which is defined as Sequence[float]. However, the code also works with passing a single int. Would it make sense to define Scalar = float | Sequence[float]? The same applies to color argument to the cv2.circle, and probably others.
• cv2.filter2D should also have at least slightly more specific type annotations.

The list of functions is quite arbitrary, I have only listed what is reported in our current code base.

Does the OpenCV Python API actually support arrays that contain np.generic but not np.number? These could be Booleans or any opaque data that does not allow arithmetic.

For functions that only support integer and floating point numbers (but not complex floats), np.ndarray[Any, np.dtype[np.integer | np.floating]] could be a better choice than np.ndarray[Any, np.dtype[np.number]]. The latter includes complex numbers, which do not support %.

from opencv.

We get tons of new Mypy errors on OpenCV 4.9 on the following functions:

• transform
• bitwise_and
• approxPolyDP
• connectedComponents
• putText
• rectangle
• etc.

Of the form:

error: Incompatible types in assignment (expression has type "ndarray[Any, dtype[generic]]", variable has type "ndarray[Any, dtype[signedinteger[_64Bit]]]")  [assignment]

Variables passed into these functions are type annotated like:

input: NDArray[np.int_] = ...

Environment

• opencv-python-headless 4.9.0.80
• mypy 1.8.0
• Pythom 3.10

[tool.mypy]
ignore_missing_imports = true
strict = true
disallow_subclassing_any = false
plugins = [
    "pydantic.mypy",
    "numpy.typing.mypy_plugin",
]

from opencv.

Right, for these types of issues, functions that preserve the input array shape and data type would need to use a TypeVar rather than independent cv2.typing.MatLike. An example of the to-be-generated code would look like this:

NumPyArrayNumeric = np.ndarray[Any, np.dtype[np.number]]
ArrayT = TypeVar("ArrayT", bound=NumPyArrayNumeric)

def erode(
    src: ArrayT,
    kernel: cv2.typing.MatLike,
    dst: ArrayT | None = ...,
    anchor: cv2.typing.Point = ...,
    iterations: int = ...,
    borderType: int = ...,
    borderValue: cv2.typing.Scalar = ...,
) -> ArrayT:
    ...

Whether the type of src is preserved depends on the individual function, though. For functions like cv2.filter2D, the ddepth argument can only be properly accounted for using overloads (if that level of typing is desired at all). Ideally, this kind of information would be encoded in the C++ documentation comments, but I don't know how to best do this with doxygen.

from opencv.

@VadimLevin Could you take a look on this?

from opencv.