Currently Flowty is built atop of OpenCV 4 which has a serious performance regression in (at least) TVL1 optical flow which makes it inpractical to use. I suggest using my other optical flow tool furnari-flow until this has been resolved. I have a benchmark repository demoing this issue and have filed a bug report. Until that is resolved I suggest you try out multiple tools and pick which ever you find most performant.

Flowty is the swiss army knife of computing optical flow. Flowty is...

• Performant—leveraging CUDA accelerated optical flow implementations.
• Easy to use—packaged in docker so you don't have to compile OpenCV and Flowty yourself.

Visit https://flowty.rtfd.org to learn more about how to obtain and use flowty. In a nutshell:

$ ls /path/to/media
video.mp4

$ docker run --rm --runtime=nvidia willprice/flowty

[Flowty help description]
  
$ docker run -it --rm \
  --runtime=nvidia \
  --mount type=bind,source=/path/to/media,target=/data \
  willprice/flowty tvl1 /data/video.mp4 /data/flow/{axis}/{index:05d}.jpg --cuda