Kinetics-700 Downsampled 112x112 Archives

This repository contains a compressed, downsampled archive version of Kinetics-700 prepared for easier local download and extraction. Instead of publishing hundreds of thousands of individual video files, the dataset is provided as split .zip archives.

The files were produced with the Kinetics-700 Dataset Downloader & Manager from the source dataset atalaydenknalbant/Kinetics-700.

Dataset Summary

Kinetics-700 is a large-scale human action recognition dataset based on short YouTube video clips. The original dataset contains approximately 700 action classes and around 635k clips, with train, validation, and test splits.

This release is a lightweight processed variant:

• Spatial resolution: 112x112
• Format: zip archives containing downsampled video files and source metadata
• Packaged archive size: approximately 9.29 GB
• Intended use: research, video classification, action recognition, and efficient experimentation on lower-storage systems

Files

File	Compressed size	Packaged file entries	Notes
train.zip	6.98 GB	536,489	Downsampled training split
validation.zip	444.91 MB	33,966	Downsampled validation split
test.zip	1.85 GB	64,533	Downsampled test split
raw.zip	14.94 MB	4	Metadata/source files used by the downloader

Usage

Download the archive files you need, then extract them locally:

unzip train.zip
unzip validation.zip
unzip test.zip
unzip raw.zip

On Windows, you can also extract the archives with File Explorer, 7-Zip, or PowerShell:

Expand-Archive train.zip -DestinationPath train
Expand-Archive validation.zip -DestinationPath validation
Expand-Archive test.zip -DestinationPath test
Expand-Archive raw.zip -DestinationPath raw

After extraction, use the split folders directly in your video classification or action recognition pipeline.

Processing Method

The downloader reduces the original large video footprint by applying a Temporal Segment Networks style preprocessing workflow:

1. Downscale clips to 112x112.
2. Sample a compact temporal sequence from each clip.
3. Save only the processed lightweight files.
4. Package each split into a zip archive for easier distribution.

This makes the dataset much smaller than the original Kinetics-700 release while preserving a trainable action-recognition structure.

Source Dataset Details

The upstream Hugging Face dataset card for atalaydenknalbant/Kinetics-700 describes Kinetics-700 as:

• 700 human action classes
• Approximately 635k total video clips
• Approximately 10 seconds per clip
• Training, validation, and test splits
• YouTube-sourced video clips with timestamp annotations

The upstream card lists the raw dataset as roughly 930 GB on Hugging Face and notes copyright, YouTube terms, demographic/source bias, and research-use considerations.

Limitations

• This is a downsampled variant, not the original full-resolution Kinetics-700 dataset.
• Some source videos may be unavailable, removed, private, corrupted, or otherwise impossible to process.
• The dataset inherits the limitations of YouTube-sourced data, including geographic, demographic, cultural, and platform biases.
• Test labels may be withheld or incomplete depending on the original source metadata.
• Users are responsible for complying with YouTube's Terms of Service, applicable copyright rules, and the original dataset usage terms.

Citation and Attribution

Please cite and acknowledge the original Kinetics dataset creators and the upstream Hugging Face source dataset when using this processed release:

• Upstream dataset: atalaydenknalbant/Kinetics-700
• Downloader/processing tool: Kinetics-700 Dataset Downloader & Manager