Kinetics#

Format specification#

Kinetics 400/600/700 is a video datasets for action recognition task. Dataset is available for downloading here

Supported media type:

  • Video

Supported type of annotations:

  • Label

Supported attributes for labels:

  • time_start (integer) - time (in seconds) of the start of recognized action

  • time_end (integer) - time (in seconds) of the end of recognized action

Import Kinetics dataset#

A Datumaro project with a Kinetics dataset can be created in the following way using CLI:

datum project create
datum project import -f kinetics <path_to_dataset>

Or using Python API:

import datumaro as dm

dataset = dm.Dataset.import_from('<path_to_dataset>', format='kinetics')
├── test.csv
├── train.json
├── train
│   ├── <name_of_video_1_with_yt_id>.avi # extension of video could be other
│   ├── <name_of_video_2_with_yt_id>.avi
│   ├── ...
└── test
    ├── <name_of_video_100_with_yt_id>.avi # extension of video could be other
    ├── <name_of_video_101_with_yt_id>.avi
    ├── ...

Kinetics dataset has two equivalent annotation file formats: .csv and .json. Datumaro supports both, but in case when two annotation files have same names but different extensions Datumaro will use .csv.

Note: name of each video file must contain youtube_id of this video, that specified in annotation file. And to speed up the import, you can leave only the youtube_id in the video filename.

See the full list of supported video extensions here.