Running feature extraction on own videos.

[Andrew-Zhang]

Hello!

I am trying to extract the features of my own videos using Slowfast. Do you recommend using feature_extractor.py, or is there a better way of extracting these features from new videos? Thanks!

[TheShadow29]

Hi @Andrew-Zhang

Sorry for the delayed reply. Yes, you can use feature_extractor.py for any videos. But it needs the videos to be extracted into image frames. Alternatively, you can just decode them during the feature extraction process using something like ffmpeg or decord.

You can also check out pytorchvideo if you want to use more recent backbones.

Let me know if that answers your questions.

[minuenergy]

Hi @Andrew-Zhang

Sorry for the delayed reply. Yes, you can use feature_extractor.py for any videos. But it needs the videos to be extracted into image frames. Alternatively, you can just decode them during the feature extraction process using something like ffmpeg or decord.

You can also check out pytorchvideo if you want to use more recent backbones.

Let me know if that answers your questions.

I want to reproduce your jobs

And I wonder what model's feature can get (15, 2048) shape features?
and also ( 15, 4 ) features either !

I try i3d which comes (5, 2048),,
please tell me which feature extractor models should i use
Thank you !!

[TheShadow29]

@minuenergy Sorry for the delayed reply.

You can find the feature extraction code here: https://github.com/TheShadow29/VidSitu/blob/main/vidsitu_code/feat_extractor.py

See the instructions here: https://github.com/TheShadow29/VidSitu/blob/main/data/DATA_PREP.md