Do you pre-extract long-term features and store the features on disk?

[coolbay]

Do you pre-extract features of objects and store the features on disk? I cannot find any operation in data processing or training steps to store the pre-extracted object features from the long-term sequence. Do you extract the object features online? If so, it still takes a large memory. I am very confused about this.

Thank you very much!

[chaoyuaw]

Among the datasets we released, we only use "object" features for "EPIC-Kitchens-Noun". We extracted those features using Detectron and store them on disk. We load them in memory when using our LFB-model. The code for object feature extraction isn't part of this repository, because it involves only standard feature extraction without modification of Detectron. We provide the yaml and model weights of the object detector at
https://github.com/facebookresearch/video-long-term-feature-banks/blob/master/GETTING_STARTED.md#epic-kitchens-noun-lfb

The extracted features are available at
https://github.com/facebookresearch/video-long-term-feature-banks/blob/master/DATASET.md#epic-kitchens