Dataset Preparation

February 28, 2024 ยท View on GitHub

We provide our labels in the labels directory.

Kinetics

  1. We download the official version of Kinetics from the dataset provider and videos are resized using code here.

  2. After all the videos are downloaded, prepare the csv files for training and validation as train.csv and val.csv. The format of the CSV file is:

<path_1>,<label_1>
<path_2>,<label_2>
...
<path_n>,<label_n>

where <path_i> points to a video file, and <label_i> is an integer between 0 and num_classes - 1.

Since some videos may no longer be available, it will lead to a small performance gap.

Something-Something-v2

  1. Please download the dataset and annotations from the dataset provider.

  2. Set up the training and validation CSV files as above for Kinetics.

Diving-48

  1. Please download the dataset and annotations from the dataset provider. Note that we use the V2 splits.

  2. Set up the training and validation CSV files as above for Kinetics.

ActivityNet-v1.3

  1. Please request access for the full v1.3 dataset here. Download the entire dataset including all missing videos.

  2. Download the annotations JSON file here

  3. Set up the training and validation CSV files as above for Kinetics.