Feature Extraction

We have used the following scripts to extract features for the collected data.

These primarily include video recognition models and multimodal understanding models.

Our feature primarily include two categories:

1. Segment features
2. Frame features

Segment Features

We have used the following models to extract segment features:

1. SlowFast
2. 3DResNet
3. Omnivore
4. X3D
5. Imagebind

Frame Features

1. Omnivore
2. Imagebind
3. TSM

Multimodal Understanding Features

1. Depth - Imagebind
2. Audio - Imagebind
3. Text - (Video --> Lavila --> Imagebind)

How to extract features

Run scripts

1. You can download the data using the downloader script
2. Place the data in folders and change the respective paths in the scripts

Use extracted features

1. You can download all the extracted features used to train all our models from here