Question about feature extraction pipeline

[jun297]

Hi, I am trying to extract features from custom video clips with TSN.
Each video clip has about 5 min length

I refered code & readme in tools/data/THUMOS14 or activitynet
What I understand is that tools/misc/clip_feature_extraction.py is the script for extraction from a given config.

What make me confusing is that: to extract feature from ActivityNet or THUMOS, first it requires to extract frames from them, and make features from the extracted frames.

However, https://github.com/open-mmlab/mmaction2/blob/0.x/docs/en/feature_extraction.md does not extract frame.
it directly extract features from video.

My question is: when is extracting frames required? what is the difference?

Thank you in advance.

[jun297]

To extract feature, you can refer regonizer2D backbone