Document OpenCV 4.9 Python changes behavior with readonly numpy arrays

[vrabaud]

System Information

OpenCV 4.x HEAD, Linux, Python 3.11

Detailed description

When some functions (e.g. cv::rectangle, cv::circle) are called on a readonly argument in OpenCV 4.8.0, they do modify it. It is unclear with the doc if the input should also be modified or not as the function returns an image, cf https://docs.opencv.org/4.8.0/d6/d6e/group__imgproc__draw.html#ga07d2f74cadcf8e305e810ce8eed13bc9.

Still, In OpenCV HEAD in 4.x, we get the following error:

error: OpenCV(4.8.0-dev) :-1: error: (-5:Bad argument) in function 'rectangle'
> Overload resolution failed:
>  - img marked as output argument, but provided NumPy array marked as readonly
>  - Expected Ptr<cv::UMat> for argument 'img'
>  - img marked as output argument, but provided NumPy array marked as readonly
>  - Expected Ptr<cv::UMat> for argument 'img'

I guess the correct behavior is to always make a copy of the image before calling cv::rectangle and return it?

Steps to reproduce

import cv2
from PIL import Image
import numpy as np
img=np.asarray(Image.open("/tmp/img.png"))
print(img.flags)
cv2.rectangle(img,0,1,1,1)

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[asmorkalov]

Related patch: #24026
Original issue: opencv/opencv-python#859

In case if the numpy array was created by external package write access to RO array may case segmentation fault. Please see details in related OpenCV-Python issue.

[vrabaud]

Thx for the pointers, I missed them. I am totally ok with the new behavior, it is a fix.

I had to go through https://docs.opencv.org/4.8.0/da/d49/tutorial_py_bindings_basics.html to get that Header parser can understand the input and output arguments from keywords like InputArray, OutputArray etc.. Maybe a little bit of text could be added for users (not sure if that is the right page, as it is a bit technical). Something like: The InputArray, OutputArray, InputOutputArray semantics are kept in Python: anything that is modified in C++ will be modified in Python. In addition, Python bindings also output anything that is an OutputArray, InputOutputArray or marked as CV_OUT.

[bitsdaddy]

.astype(np.uint8)

using this with the image solved the problem for me