Kinetics 700: YouTube Video Clips for Human Action Recognition

The Kinetics-700 dataset is a large-scale collection of YouTube video URLs for human action recognition, containing 700 distinct action classes. It is an extension of the original Kinetics dataset, with each video clip being approximately 10 seconds long and depicting a single human action. The dataset was uploaded by author 'bitmind' to the Hugging Face platform, with a last recorded update in October 2025.

Use Cases

• Training action recognition models based on the 700 diverse human action classes.
• Benchmarking video classification model performance on a large-scale dataset.
• Developing models for complex and diverse human actions as described in the dataset's purpose.
• Fine-tuning pre-trained models on specific subsets of the 700 action categories.

Strengths

• Contains 700 distinct human action classes, providing a wide range of categories.
• Each video clip is approximately 10 seconds long, offering a consistent temporal format.
• Described as a large-scale collection, suggesting a substantial number of video clips.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

YouTube video URLs.

Collection Method

Collection of publicly available video URLs, likely curated for action recognition.

Freshness

Last updated 2025-10-26 01:25:48; freshness should be verified.