Denseflow seems incredibly slow

[btalberg]

I'm preparing the AVA dataset for training the SlowFast network, and the step that runs denseflow seems to be running incredibly slow. Here's a snippet from the output of bash extract_frames.sh:

1 videos (21603 frames, 21602 tvl1 flows) processed, using 10812.7s, decoding speed 1.99792fps, flow speed 1.99783fps

We're running a GeForce RTX 2080 SUPER with 8GB of memory. Our CUDA version is 11.4.

Are 2 FPS normal for denseflow? Any advice on how to speed this up?

And FWIW, the step that extracts RGB frames seems to be running very quickly. Here's another snippet from the log file:

1 videos (22525 frames, 0 tvl1 flows) processed, using 94.469s, decoding speed 238.438fps, flow speed 0fps

Thanks for your help!

(One additional comment/question, are the optical flow images actually necessary for SlowFast? Generating the RGB images appears to be required, but could I skip the step of generating flow output?)