Notes:
(1) Requirements OpenCV-4.4.0 with DNN support needed.
-DWITH_CUDA=ON -DWITH_CUDNN=ON -DOPENCV_DNN_CUDA=ON -DENABLE_FAST_MATH=1 -DCUDA_FAST_MATH=1 -DCUDA_ARCH_BIN=${CUDA_ARCH_BIN} -DWITH_CUBLAS=1
where the ${CUDA_ARCH_BIN} should be the compute capability of the specific gpu which can be found in the wiki.

(2) The input size is currently limited to a fixed size of 448x320.

(3) The speed is around 4 FPS for a single 130-frame video stream. As a comparison, TV-L1 is around 40 FPS (320x240).