joaanna/something_else

Is there source code for frame extraction released?

meifish opened this issue · 6 comments

meifish commented

Hi author,

This is an impressive study and work! I really like the concept of compositional object interaction and the methodology.
I have a question regarding the frames extraction.

While trying to apply the annotate_videos.py on the video frames, I noticed that the number of frames I use cv2.VideoCapture(..) to extract is inconsistent with the frames in your demo video folders, so as to the frames number specified in annotation JSON file. It is very close but there's always a frame or two short from your version. I wonder is there frame extraction source code could be released, so to bridging the data consistency? Thank you!!

patrolli commented

@meifish ，hello, did you get the solution of the frame extraction?

meifish commented
patrolli commented

@meifish Could you please share a copy of this preprocessing code? I also tried to use ffmpeg to extract the frames, but I am not sure about the final size of the extracted frame files.

meifish commented

Hello @patroli:

If your extracted frames matched their demo video here, then it should be correct:
https://github.com/joaanna/something_else/tree/master/videos

I used the script here, search "extract frames" in the page, and change the number 30 to 12 (i.e., fps=12). Also in the out_name variable, change the %6d to %4d (i.e., 000012.jpg versus 0012.jpg):
https://github.com/facebookresearch/SlowFast/blob/master/slowfast/datasets/DATASET.md#something-something-v2

patrolli commented

Thanks for your enthusiastic help!

meifish commented

No problem! Let's mark it as solved now.