The Kinetics Human Action Video Dataset
A large dataset of 400 human actions in short video clips, collected from YouTube, to help train computer vision models to recognize human activities.
Why it matters
This large-scale dataset of human actions provides a valuable resource for developing and benchmarking AI models that can recognize and understand human activities.
Key Points
- 1The dataset contains 400 different human actions
- 2Each action has at least 400 short video clips of around 10 seconds
- 3The clips are from diverse real-world YouTube videos to provide natural examples
- 4The dataset is meant to be a practical tool for training AI models to recognize human movements
Details
The Kinetics Human Action Video Dataset is a large collection of short video clips focused on real people performing various actions and activities. The dataset contains 400 different human actions, with at least 400 video clips per action, each around 10 seconds long. The clips are sourced from diverse YouTube videos to provide natural and varied examples of human movements and behaviors. The dataset is intended to be a practical tool for training computer vision and activity recognition models, allowing them to learn from a broad range of human actions seen in daily life. The creators also tested the dataset for potential biases that could arise from uneven distribution of examples across actions, to ensure the data is suitable for robust model training.
No comments yet
Be the first to comment