OpenCV conflict

[VieiraJardel]

I compiled OpenCV 4.5.4 on my Jetson Nano, but during the 'python3 -m pip install --no-deps dist/mediapipe-0.8.9_cuda102-cp36-cp36m-linux_aarch64.whl' step, version 4.8.0 is being installed. I don't want to install this newer version because it results in a loss of GPU processing. Could you recommend something to address this issue of version conflict during installation?

[anion0278]

I have provided the commands for installation here in readme, have you tried them?

especially pip3 install protobuf==3.19.4 opencv-python==4.5.3.56 dataclasses mediapipe-0.8.9_cuda102-cp36-linux_aarch64.whl

[VieiraJardel]

I followed your commands and tried not to select OpenCV, but it didn't work. When I installed MediaPipe, OpenCV 4.8.1 was included automatically.

Is it possible to avoid installing OpenCV during the build and instead use dependencies from another installation, such as a custom-built version in my case?

[anion0278]

well, this command should have done exactly what you needed. Unfortunately, I don't have access to Jetson anymore to try to figure out other ways.
Nevertheless, I guess by

loss of GPU processing

you mean that your OpenCV 4.5.4 is compiled with CUDA support. This is a common misconception that compiling OpenCV with CUDA makes all operations faster by running them on GPU. To actually use GPU you would still need to modify your code and actually upload data to-and-from GPU, see tutorial. So unless your code is actually adapted to do that, you are not losing anything by installing a new non-cuda version of OpenCV.

[VieiraJardel]

Hi @anion0278 !
Can you please provide the link again? It's currently not working

[anion0278]

Hi,
well, the link is now completely dead (removed), but you can find a similar tutorial here.