Video Dataset

This guide will show you how to configure your dataset repository with video files.

A dataset with a supported structure and file formats automatically has a Dataset Viewer on its page on the Hub.

Additional information about your videos - such as captions or bounding boxes for object detection - is automatically loaded as long as you include this information in a metadata file (metadata.csv/metadata.jsonl/metadata.parquet).

Alternatively, videos can be in Parquet files or in TAR archives following the WebDataset format.

Only videos

If your dataset only consists of one column with videos, you can simply store your video files at the root:

my_dataset_repository/
├── 1.mp4
├── 2.mp4
├── 3.mp4
└── 4.mp4

or in a subdirectory:

my_dataset_repository/
└── videos
    ├── 1.mp4
    ├── 2.mp4
    ├── 3.mp4
    └── 4.mp4

Multiple formats are supported at the same time, including MP4, MOV and AVI.

my_dataset_repository/
└── videos
    ├── 1.mp4
    ├── 2.mov
    └── 3.avi

If you have several splits, you can put your videos into directories named accordingly:

my_dataset_repository/
├── train
│   ├── 1.mp4
│   └── 2.mp4
└── test
    ├── 3.mp4
    └── 4.mp4

See File names and splits for more information and other ways to organize data by splits.

Additional columns

If there is additional information you’d like to include about your dataset, like text captions or bounding boxes, add it as a metadata.csv file in your repository. This lets you quickly create datasets for different computer vision tasks like video generation or object detection.

my_dataset_repository/
└── train
    ├── 1.mp4
    ├── 2.mp4
    ├── 3.mp4
    ├── 4.mp4
    └── metadata.csv

Your metadata.csv file must have a file_name column which links video files with their metadata:

file_name,text
1.mp4,an animation of a green pokemon with red eyes
2.mp4,a short video of a green and yellow toy with a red nose
3.mp4,a red and white ball shows an angry look on its face
4.mp4,a cartoon ball is smiling

You can also use a JSONL file metadata.jsonl:

{"file_name": "1.mp4","text": "an animation of a green pokemon with red eyes"}
{"file_name": "2.mp4","text": "a short video of a green and yellow toy with a red nose"}
{"file_name": "3.mp4","text": "a red and white ball shows an angry look on its face"}
{"file_name": "4.mp4","text": "a cartoon ball is smiling"}

And for bigger datasets or if you are interested in advanced data retrieval features, you can use a Parquet file metadata.parquet.

Relative paths

Metadata file must be located either in the same directory with the videos it is linked to, or in any parent directory, like in this example:

my_dataset_repository/
└── train
    ├── videos
    │   ├── 1.mp4
    │   ├── 2.mp4
    │   ├── 3.mp4
    │   └── 4.mp4
    └── metadata.csv

In this case, the file_name column must be a full relative path to the videos, not just the filename:

file_name,text
videos/1.mp4,an animation of a green pokemon with red eyes
videos/2.mp4,a short video of a green and yellow toy with a red nose
videos/3.mp4,a red and white ball shows an angry look on its face
videos/4.mp4,a cartoon ball is smiling

Metadata files cannot be put in subdirectories of a directory with the videos.

More generally, any column named file_name or *_file_name should contain the full relative path to the videos.

Video classification

For video classification datasets, you can also use a simple setup: use directories to name the video classes. Store your video files in a directory structure like:

my_dataset_repository/
├── green
│   ├── 1.mp4
│   └── 2.mp4
└── red
    ├── 3.mp4
    └── 4.mp4

The dataset created with this structure contains two columns: video and label (with values green and red).

You can also provide multiple splits. To do so, your dataset directory should have the following structure (see File names and splits for more information):

my_dataset_repository/
├── test
│   ├── green
│   │   └── 2.mp4
│   └── red
│       └── 4.mp4
└── train
    ├── green
    │   └── 1.mp4
    └── red
        └── 3.mp4

You can disable this automatic addition of the label column in the YAML configuration. If your directory names have no special meaning, set drop_labels: true in the README header:

configs:
  - config_name: default  # Name of the dataset subset, if applicable.
    drop_labels: true

Large scale datasets

WebDataset format

The WebDataset format is well suited for large scale video datasets. It consists of TAR archives containing videos and their metadata and is optimized for streaming. It is useful if you have a large number of videos and to get streaming data loaders for large scale training.

my_dataset_repository/
├── train-0000.tar
├── train-0001.tar
├── ...
└── train-1023.tar

To make a WebDataset TAR archive, create a directory containing the videos and metadata files to be archived and create the TAR archive using e.g. the tar command. The usual size per archive is generally around 1GB. Make sure each video and metadata pair share the same file prefix, for example:

train-0000/
├── 000.mp4
├── 000.json
├── 001.mp4
├── 001.json
├── ...
├── 999.mp4
└── 999.json

Note that for user convenience and to enable the Dataset Viewer, every dataset hosted in the Hub is automatically converted to Parquet format up to 5GB. Since videos can be quite large, the URLs to the videos are stored in the converted Parquet data without the video bytes themselves. Read more about it in the Parquet format documentation.