Kinetics-700
Google DeepMind (hosted by the Common Visual Data Foundation)
650,000 ten-second YouTube clips of people doing 700 things, from yoga to changing a tire.
Human-action video clips, each about 10 seconds long and labeled with one of 400, 600, or 700 action classes depending on version. It is the main benchmark and pretraining set for video recognition. Because the source videos live on YouTube, many links died over time, so CVDF now hosts copies of Kinetics-400 and Kinetics-700-2020.
Access
- Cost
- FreeFree download scripts for hosted copies.
- License
- License not statedNot stated on the hosting page; videos remain YouTube content owned by uploaders
- Sign-up
- No sign-up
- How you get it
- Bulk downloadMP4, CSV
What's in it
- One row is
- One roughly 10-second video clip with one action label
- Covers
- YouTube videos worldwide
- Years
- Kinetics-400 released 2017; Kinetics-700-2020 in 2020
- Updated
- One-time release
- Size
- Up to about 650,000 clips across 700 classes, with at least 700 clips per class
Quality
The standard video benchmark with human-verified labels; versions overlap, and mixing them leaks test data.
- Kinetics-400 validation clips largely appear in Kinetics-700 training data
- Original YouTube links decay over time
- Uploaders did not consent to research use
Connects to
Questions it could answer
- Which actions are hardest to tell apart from a single frame?
- What share of the original YouTube clips are still online?