findContours hangs in subprocess

[mbway]

findContours can only be run in worker processes until the main process runs findContours, after which point if any worker process tries to call the function again, it hangs indefinitely. This only applies to processes which are forked from the main process.

For example the following code snippet:

import multiprocessing
import cv2 as cv
import numpy as np

def get_contours(image: np.ndarray, name: str):
    print(f'before {name}')
    print(cv.findContours(image.copy(), mode=cv.RETR_CCOMP, method=cv.CHAIN_APPROX_SIMPLE))
    print(f'after {name}')

if __name__ == '__main__':
    multiprocessing.set_start_method('fork')

    image = np.zeros((1000, 1000), dtype=np.uint8)
    image[1, 1] = 1

    p = multiprocessing.Process(target=get_contours, args=(image, 'parallel 1'))
    p.start()
    p.join()

    p = multiprocessing.Process(target=get_contours, args=(image, 'parallel 2'))
    p.start()
    p.join()

    get_contours(image, 'serial 1')
    get_contours(image, 'serial 2')

    p = multiprocessing.Process(target=get_contours, args=(image, 'parallel 3'))
    p.start()
    p.join()

    print('finished')

The output is

before parallel 1
([array([[[1, 1]]], dtype=int32)], array([[[-1, -1, -1, -1]]], dtype=int32))
after parallel 1
before parallel 2
([array([[[1, 1]]], dtype=int32)], array([[[-1, -1, -1, -1]]], dtype=int32))
after parallel 2
before serial 1
([array([[[1, 1]]], dtype=int32)], array([[[-1, -1, -1, -1]]], dtype=int32))
after serial 1
before serial 2
([array([[[1, 1]]], dtype=int32)], array([[[-1, -1, -1, -1]]], dtype=int32))
after serial 2
before parallel 3

then it hangs

Notes

• if multiprocessing.set_start_method('spawn') is used, or threads are used instead then the program does not hang.
• if the image is sufficiently small (e.g. (10, 10)) then even if fork is used, the program does not hang (on my machine at least)

Expected behaviour

findContours is called in the main process and each worker process then exits.

Actual behaviour

After findContours has been run from the main process, at the next call to findContours in a worker process, the program hangs

Steps to reproduce

$ lsb_release -a
No LSB modules are available.
Distributor ID:	Ubuntu
Description:	Ubuntu 18.04.3 LTS
Release:	18.04
Codename:	bionic
$ python
Python 3.7.3 | packaged by conda-forge | (default, Jul  1 2019, 21:52:21) 
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2; cv2.__version__
'4.1.1'
>>> import numpy as np; np.__version__
'1.16.4'
$ conda list|grep opencv
opencv                    4.1.1            py37ha799480_1

[skvark]

This is a known issue in OpenCV: opencv/opencv#5150

In general, avoid multiprocessing with OpenCV.

There's nothing I can do about OpenCV bugs in this repo. This repository just provides automated toolchain to build opencv-python wheels.

Also, you seem to be using Conda's OpenCV package (repo: https://github.com/conda-forge/opencv-feedstock, site: https://anaconda.org/conda-forge/opencv). This package (https://pypi.org/project/opencv-python/) is a PyPI package and can be obtained via pip install opencv-python.

[mbway]

I suspected it might be an issue with opencv rather than the python wrapper but I thought I'd check here first. Thanks for the fast response.