Feature Extraction

[Camillo4eyes]

Hi, can you provide more informations about the feature extraction? I would like to use this fantastic model on my dataset but I don't know how to extract the features to feed to the encoder.

[ChinaYi]

We used the pretrained I3D model on Kinetics. Then we used a window of size 21 around each frame, i.e. [t-10, t+10] to extract frame-wise features. For more details about the features please check this repo https://github.com/ahsaniqbal/Kinetics-FeatureExtractor

[Camillo4eyes]

So, considering a stack size (or window size) of 21, the extracted features will have size (n_frames / 21) x 1024, right? (which then, by concatenating the stram RGB and flow will become (n_frames / 64) * 2048). Didn't you use any step between one window and another? Sorry for the many questions, but I'm really curious and I really want to use this model!

[ChinaYi]

Sorry for my late response :) really got some busy days. We do not use any step between windows since we have to collect features per-frame with full fps rate.