Features extraction and preparation of new video dataset

[MH-Python]

We want to use your pre-trained models (K700) to exctract visual features of some videos and then use these features in a classifier.
When we checked the last layer output (before the classification layer of 700), it contained a 1024 1D vector for each 3, 16, 224, 224 input. Can this be cosidered as the latent-features, or there is another way or a layer that containes more semantic features?

The dataset of video was prepared as short videos of 64 frames (fps 30) and then downsampled to 16 frames by a temporal step of 4,
video_frames = video_frame[::4]
doest this sound as valid processing of the video before passing it to the model for features extraction.

Thank you