OpenCV optical flow speed comparison

I've found OpenCV's GPU TVL1 implementation to be much slower in v4 than in v2. This repository serves as an example demonstrating the issue.

Set up

Ensure you have docker 19.03+ with an NVIDIA card present on your system. Build the docker images (this handles building the base OpenCV images + the optical flow demo application)

$ make flow-images

Download test media and extract frames:

$ ffmpeg -i "https://github.com/MarkAYoder/esc-media/raw/master/BigBuckBunny_640x360.m4v" -t 00:00:10 -qscale 2 frames/frame_%010d.jpg

Run speed test

Discard the first results as they will include the time spent by the nvidia driver generating binaries for the current GPU from the PTX files.

$ time docker run \
    --gpus '"device=0"' \
    --rm -it \
    -v $PWD/frames:/input \
    -v $PWD/flow/opencv2:/output \
    -v $PWD/.cache-opencv2:/cache/nv \
    willprice/furnari-flow:opencv2 

...
0.03user 0.02system 0:14.57elapsed 0%CPU (0avgtext+0avgdata 63544maxresident)k
0inputs+0outputs (0major+7956minor)pagefaults 0swaps

$ time docker run \
    --gpus '"device=0"' \
    --rm -it \
    -v $PWD/frames:/input \
    -v $PWD/flow/opencv4:/output \
    -v $PWD/.cache-opencv4:/cache/nv \
    willprice/furnari-flow:opencv4

...
0.04user 0.02system 2:31.88elapsed 0%CPU (0avgtext+0avgdata 63404maxresident)k
0inputs+0outputs (0major+7877minor)pagefaults 0swaps