opencv-python-headless depends on windows media feature pack after 4.7 update

[matteoaletti]

Expected behaviour

On previous version (4.6.0.66) of opencv-python-headless, the package did not depend on windows media feature pack.

>> pyi-bindepend.exe "C:\Users\user\anaconda3\envs\cv246\Lib\site-packages\cv2\cv2.pyd"
C:\Users\user\anaconda3\envs\cv246\Lib\site-packages\cv2\cv2.pyd {'WSOCK32.dll', 'ole32.dll', 'KERNEL32.dll', 'OLEAUT32.dll', 'python3.dll'}

Actual behaviour

More recently (4.7.0.68) of opencv-python-headless, the package depends on windows media feature pack that is not available in windows docker image (windows/servercore:ltsc2019) .

>> pyi-bindepend.exe "C:\Users\user\anaconda3\envs\cv247\Lib\site-packages\cv2\cv2.pyd"
C:\Users\user\anaconda3\envs\cv247\Lib\site-packages\cv2\cv2.pyd {'MFReadWrite.dll', 'MF.dll', 'MFPlat.DLL', 'OLEAUT32.dll', 'KERNEL32.dll', 'SHLWAPI.dll', 'ole32.dll', 'WSOCK32.dll', 'python3.dll'}

Steps to reproduce

I have installed opencv-python-headless in a conda python 3.9 environment with

pip install opencv-python-headless

[asmorkalov]

@matteoaletti Thanks for the issue report. I have not found explicit changes in setup.py or build scripts that manage Media Foundation dependency. Looks like it's some side effect of Python 3.6 -> 3.7 or Github actions environment update. Could you list other container specific issues here. I want to discuss the next steps with colleagues.

[matteoaletti]

Thank you for your answer.

I am not sure that there are any other container specific issues. We are using a conda environment with python 3.9 inside a windows docker container (windows/servercore:ltsc2019) and when we use the opencv-python-headless 4.6.0.66 we can import cv2. Same environment with version 4.7.0.68 and our tests produces an error during the import of cv2:

ImportError: DLL load failed while importing cv2: The specified module could not be found

We tried to debug it ourselves and found out the new DLL dependencies mentioned above.
Everything works fine outside the container in a normal windows OS where those libraries are available.

For now, we simply pinned the version to the latest 4.6.

[alalek]

@asmorkalov Need to check G-API module and/or obsensor feature.

[yoonjs2]

Any news to resolve this issue after 4.8 version? I encountered this issue when import cv2 in aws codebuild docker container image (windows server 2019) . I tried 4.6 4.7 4.8 but only 4.6 works.

[asmorkalov]

Decided to switch to plugins for videoio in headless package on Windows. In case of good luck, it'll be part of 4.9.0.x.

[matteoaletti]

@asmorkalov I see that a new release of opencv is out, was this switch included in 4.9.0.80 ?

I still see the same dependencies (and errors) as for versions 4.7 and 4.8

[Avasam]

Likely duplicated by #856 (although that issue is more popular)

[pwuertz]

The current opencv-python-headless 4.11 package still depends on Media Foundation Platform libraries:

pyi-bindepend cv2.pyd
  KERNEL32.dll => C:\Windows\system32\KERNEL32.dll
  MF.dll => C:\Windows\system32\MF.dll
  MFPlat.DLL => C:\Windows\system32\MFPlat.DLL
  MFReadWrite.dll => C:\Windows\system32\MFReadWrite.dll
  OLEAUT32.dll => C:\Windows\system32\OLEAUT32.dll
  SHLWAPI.dll => C:\Windows\system32\SHLWAPI.dll
  WSOCK32.dll => C:\Windows\system32\WSOCK32.dll
  ole32.dll => C:\Windows\system32\ole32.dll
  python3.dll => C:\ProgramData\scoop\apps\python\current\python3.dll

Importing the cv2 module remains unsupported within Windows Server Core containers or Windows N systems, where Media Foundation is not available.

However, the Media Foundation libraries are present in the mcr.microsoft.com/windows/server:ltsc2025 Docker image. If you are currently blocked by this issue, switching to that base image may be a viable workaround. In my testing, importing cv2 succeeds in this environment (though I haven't yet validated more complex functionality).

[asmorkalov]

Related ticket: #263
Will try to make MSMF plugin.

[avdivan]

Hi, dependencies found with MSMF

x86_64-w64-mingw32-objdump -x opencv_unpack4/cv2/cv2.pyd | grep -i media
	034fe600  <none>  0005  MFCreateSourceReaderFromMediaSource

strings opencv_unpack4/cv2/cv2.pyd | grep -iE 'mfplat|mfreadwrite|mf\.dll|evr'
MFPlat.DLL
MF.dll
MFReadWrite.dll

and
gapi/src/streaming/onevpl/demux/async_mfp_demux_data_provider.cpp:447