Drawing into a numpy view with a negative step in the channel dimension doesn't work

[tailsu]

System information (version)

• OpenCV => opencv-contrib-python-headless==4.1.1.26
• Operating System / Platform => OS X

Detailed description

I create three 3-channel images and draw a circle into them. In the second case I use the step trick A[:, :, ::-1] to flip the R and B channels in the image, emulating cv.cvtColor with COLOR_BGR2RGB. I then try to draw into that array. The result, as seen in the screenshot, is that nothing gets drawn. In the third case I explicitly make a copy of the channel-flipped array (which is actually a view) before drawing, and it works.

Expected is that cv.circle works in all three cases.

Steps to reproduce

import cv2 as cv

image1 = np.zeros((100, 100, 3), dtype='uint8')
cv.circle(image1, (50, 50), 10, (255, 255, 255), -1)

image2 = np.zeros((100, 100, 3), dtype='uint8')
image2 = image2[:, :, ::-1]
cv.circle(image2, (50, 50), 10, (255, 255, 255), -1)

image3 = np.zeros((100, 100, 3), dtype='uint8')
image3 = np.array(image3[:, :, ::-1])
cv.circle(image3, (50, 50), 10, (255, 255, 255), -1)

imshow(image1, image2, image3)

[tailsu]

imshow is a custom function that wraps ipywidgets.Image, but I tried it with cv.imwrite as well and the result is the same, so it's not a visualization issue.

[alalek]

There is lack of support in OpenCV for numpy on-the-fly reshaped arrays. OpenCV can't even copy data back if layout is not supported.

Workaround: use return value.

-cv.circle(image2, (50, 50), 10, (255, 255, 255), -1)
+image2 = cv.circle(image2, (50, 50), 10, (255, 255, 255), -1)

image1 = 128 * np.ones((100, 100, 3), dtype='uint8')
cv.circle(image1, (50, 50), 10, (255, 255, 255), -1)

image2 = 128 * np.ones((100, 100, 3), dtype='uint8')
image2 = image2[:, :, ::-1]
cv.circle(image2, (50, 50), 10, (255, 255, 255), -1)
image2_ = cv.circle(image2, (50, 50), 10, (255, 255, 255), -1)

image3 = 128 * np.ones((100, 100, 3), dtype='uint8')
image3 = np.array(image3[:, :, ::-1])
cv.circle(image3, (50, 50), 10, (255, 255, 255), -1)

cv.imshow("1", image1)
cv.imshow("2", image2)
cv.imshow("2_", image2_)
cv.imshow("3", image3)
cv.waitKey()

[alalek]

With patch from #15918 we can see this message of failed conversion:

[ WARN:0] Bindings conversion failed: Layout of the output array img is incompatible with cv::Mat (step[ndims-1] != elemsize or step[1] != elemsize*nchannels)

Unfortunately, bindings wrapper didn't fail on these arguments and did moved forward to malformed UMat overload (this is a bug).