Kinetics

Overview

Kinetics is DeepMind’s large-scale video understanding dataset containing hundreds of thousands of short clips labeled with human actions. It established the standard for video classification and pre-training video understanding models.

What’s In It

Kinetics-700 contains approximately 650,000 video clips spanning 700 human action classes. Each clip is roughly 10 seconds long, sourced from YouTube, and centered on a single human action ranging from simple to complex activities.

How It’s Used

Kinetics is the primary pre-training and evaluation dataset for video understanding models including I3D, SlowFast, and TimeSformer. Pre-trained features transfer to downstream tasks like action detection and video captioning.

Controversies

YouTube sourcing means clips disappear over time as videos are deleted, creating reproducibility challenges. Some action categories overlap or are difficult to distinguish consistently.