Kinetics400

Kinetics-400 is a large-scale, labeled video dataset designed for benchmarking action recognition in video. Released by DeepMind in 2017, it comprises 400 action classes and a substantial collection of short clips sourced from YouTube. Each video is labeled with a single action class, and the dataset is divided into training and validation splits to support model development and evaluation. Publicly released benchmarks for Kinetics-400 typically evaluate on the validation set, with some experiments using held-out test sets provided in the original research framework.

The videos in Kinetics-400 are sourced by searching for action-relevant terms and are then subjected to human annotation to verify the action label, remove duplicates, and ensure labeling quality. Clips are generally short, depicting a wide range of human activities across diverse environments, lighting, and camera viewpoints to encourage robust recognition methods.

Impact and use in research have been substantial. Kinetics-400 has served as a standard benchmark for video action recognition, helping drive advances in architectures such as 3D convolutional networks (for example, C3D and I3D), two-stream approaches, and, more recently, transformer-based models. The dataset is often used to pretrain models that are later fine-tuned on other video tasks.

Related to Kinetics-400 are larger successors, including Kinetics-600 and Kinetics-700, which expand the number of classes and clips. Access to Kinetics-400 is provided for research under terms set by the organizers, reflecting copyright considerations tied to YouTube video content.