How to use VideoMAEv2 to extract features from self-recorded videos for inference?

[iamzmchu]

Question

How do I extract features from my own recorded video using VideoMAEv2 for inference with Baseline_VideoMAEv2-Giant.ckpt?

Background

I'm trying to extract features from my own recorded video. This project seems designed for processing pre-processed video frames, but I'd like to understand:

1. How to process the raw video into the correct frame format (frame_*.jpg)

2. How to extract features from a custom video using the VideoMAEv2 model

I only know that a sliding window is needed, but I know nothing else about the details.

Thanks!

[saikat927]

Details of the feature extraction is required, otherwise the extracted feature is useless. For example I can see that "france_ligue-1/2014-2015/2015-04-05 - 22-00 Marseille 2 - 3 Paris SG" first half is a 45.14 min video and the corresponding feature has the dimension 6262x384. What is the setup to get this feature using the VideMAE model? How can I get the video feature corresponding to the match time say 2.1 min to 2.5 min?

[AugustRushG]

I have similar questions regarding the features extracted as well, im not understanding how time aligns with the feature dimension here. How does "france_ligue-1/2014-2015/2015-04-05 - 22-00 Marseille 2 - 3 Paris SG" first half which is 2714 seconds corresponds to 6262? How does this work?