cv2.imshow() not working consistently

[GokulNarayan]

Expected behaviour

imshow displays image

Actual behaviour

when calling cv2.imshow(), I mostly get a small window with the correct name but containing only a black screen. It does work sometimes however but it is unpredictable. This issue was not encountered in previous versions of opencv-contrib-python.

Steps to reproduce

• example code

import cv2
import os
import imutils

image = cv2.imread("ref_image.png")

cv2.imshow("test", image)
cv2.waitKey(0)

cv2.destroyAllWindows()

• operating system : Ubuntu 20.04
• architecture (e.g. x86): x64
• python version: 3.8
• opencv-python version: opencv-contrib-python 4.3.36

[skvark]

Yeah, probably same issue as here: #356

I have no idea what causes that behavior. Qt4 version is the same, only the base image where the wheels and Qt4 are built was changed from manylinux1 to manylinux2014.

@skvark Do you know any workaround to solve this? it happens on any version of opencv-contrib-python using ubuntu 20

[skvark]

I had a look at the linked OpenCV issue and it seems that this issue is related to Qt conflicts. There is no proper workaround for the latest wheels (see next message for some options). Best option would be to bundle Qt5 to the wheels (Qt4 is very old and does not behave well with other global Qt system installations) but there were major issues when I tried that during the manylinux version upgrade (#309). Auditwheel might as well cause the issue, it bundles the dependencies into the wheels and it might be doing something wrong.

[skvark]

How to avoid this issue until it's solved:

• downgrade to the previous release which uses manylinux1 (4.2.0 and 3.4.9)
• use some proper GUI framework to display the images such as PyQt5
• build OpenCV wheel locally and use it, instructions are in the readme: https://github.com/skvark/opencv-python#manual-builds

[skvark]

I will rebuild the manylinux2014 images and check if that fixes this issue, auditwheel has had some updates since the last build.

[skvark]

That didn't help, I'll need to probably start looking into Qt5 again...

[skvark]

Looks like I got the Qt5 builds working (at least Travis tests are passing), but I need to still check that GUI functions actually work.

Edit: still segfaulting...

[skvark]

I patched auditwheel and got Qt 5.15 working. Qt is also built in a namespace so it should not conflict with other systems Qt installations in the future. There is still some font related finetuning to do but it's far more easier to fix than the segmentation fault issue.

[skvark]

This issue should be now fixed. Fixed wheels (3.4.10.37 and 4.3.0.38) will be soon up in PyPI. It will take probably about ~24 hours for all builds to finish.

Please open a new issue if there are still problems related to GUI functions.