google-deepmind/kinetics-i3d

How should I normalize the framerate?

skaws2003 opened this issue · 2 comments

Currently I am normalizing the input video frame rate by sampling the nearest frame of the desired fps. Specifically, using fps=fps=25:round=near filter of ffmpeg.

However, I'm not sure whether this is the right way. With this way, the output TVL1 optical flow flickers since the step size for frame sampling changes.
For example, sampling a 20fps video by 25fps may result completely blank flow image for some timestep, since the kind of sampling makes redundant frames for the sampled(25fps) video.

So could you let me know the way you handled the different frame rate?
Is it just nearest sampling? or have you matched the framerate by changing the total video length?

Hi, the preprocessing has been open-sourced in the mediapipe repository. Check https://github.com/google/mediapipe/tree/master/mediapipe/examples/desktop/media_sequence

Thank you!