Revisiting OpenCV with CUDA build

[apatsekin]

Comment:

Hey guys! Revisiting this issue #109

I think things changed since 2018. Conda-forge actively maintains major CV/ML libraries delivering cuda builds for: tensorflow, pytorch, mmcv - and this is really great and super useful. What blockers do we have to go the same route with OpenCV? I'd be happy to contribute if it's a matter of human resources.

cv2.dnn is incredibly robust and efficient model runner that supports models from onnx/tensorflow/darknet/torch/caffe. Having it limited to CPU only is really a bummer.

[hmaarrfk]

PRs welcome from my end

[hmaarrfk]

try to follow the pytorch library in terms of naming

[traversaro]

Once done with #363 I am interested in working on this, even if I understand that increasing the build matrix even more may be problematic. Perhaps we could just enable it for a subset of ffmpeg/protobuf/python?