Kinetics Dataset Download ~repack~ May 2026

The CVDF hosts many versions of the dataset on AWS, which is often more reliable than crawling YouTube directly.

The Kinetics dataset, released by , is a benchmark for human action recognition in video. It consists of curated YouTube clips, each lasting approximately 10 seconds, covering diverse human-object and human-human interactions. Clips (Approx.) Key Feature Kinetics-400 The original foundational version. Kinetics-600 Expanded classes and samples. Kinetics-700 Largest standard version. Kinetics-700-2020 Replenishes deleted videos from older versions. Top Ways to Download Kinetics 1. Using Official CVDF Scripts (Recommended for Stability)

: Use the corresponding extractor script (e.g., bash ./k600_extractor.sh ) to unzip the tarballs. 2. Using FiftyOne (Easiest for Fast Prototyping) kinetics dataset download

: Use the shell scripts for your specific version (e.g., k400_downloader.sh or k700_2020_downloader.sh ).

To download the Kinetics dataset, you can use official repository scripts from CVDF (Common Visual Data Foundation) or high-level libraries like FiftyOne and MMAction2 . The CVDF hosts many versions of the dataset

The FiftyOne Dataset Zoo allows you to download specific subsets or full versions with a single Python command.

Preparing Kinetics-[400/600/700] - MMAction2's documentation! Clips (Approx

: git clone https://github.com/cvdfoundation/kinetics-dataset.git