Skip to main content

Ego4D Dataset Download CLI

The Ego4D CLI can be installed via pip and provides access to the Ego4D datasets.

Getting Started


Install via pip (conda support coming):

pip install ego4d


The CLI requires an AWS license is properly setup as per start here

Basic Usage

In your python environment, use the ego4d command line directly:

ego4d --output_directory="~/ego4d_data" --datasets full_scale annotations --metadata

(Alternatively, use traditional python module syntax: python -m ego4d.cli.cli --output_directory="~/ego4d_data" --datasets full_scale annotations --metadata)

This will download all the full scale Ego4D v1 video files and annotations to a directory on your local computer at ~/ego4d_data/v1/full_scale and ~/ego4d_data/v1/annotations, as well the master metadata file at ~/ego4d_data/v1/ego4d.json.

Note that if you want to use the AWS credentials stored in a different named profile than "ego4d", or the system default (default), you can change the aws_profile_name flag to the name of the profile that you want to use.

Detailed Flags

Flag NameDescription
--dataset[Required] A list of identifiers to download: [annotations, full_scale, clips] Each dataset will be stored in folders in the output directory with the name of the dataset (e.g. output_dir/v1/full_scale/) and manifest.
--output_directory[Required]A local path where the downloaded files and metadata will be stored
--metadata[Optional] Download the primary ego4d.json metadata at the top level (Default: True)
--benchmarks[Optional] A list of benchmarks to filter dataset downloads. One of {'AV', 'EM', 'FHO', 'MQ', 'NLQ', 'VQ', 'goalstep'}
-y --yes[Optional] If this flag is set, then the CLI will not show a prompt asking the user to confirm the download. This is so that the tool can be used as part of shell scripts.
--aws_profile_name[Optional] Defaults to “default”. Specifies the AWS profile name from ~/.aws/credentials to use for the download
--video_uids[Optional] List of video or clip UIDs to be downloaded. If not specified, all relevant UIDs will be downloaded.
--video_uid_file[Optional] Path to a whitespace delimited file that contains a list of UIDs. Mutually exclusive with the video_uids flag.
--universities[Optional] List of university IDs. If specified, only UIDs from the S3 buckets belonging to the listed universities will be downloaded.
--version[Optional] A version identifier - e.g. “v1”
--no-metadata[Optional] Bypass the ego4d.json metadata download
--config[Optional] Local path to a config JSON file. If specified, the flags will be read from this file instead of the command line


The following datasets are available (not exhaustive):

annotationsThe full set of annotations for the majority of benchmarks.
full_scaleThe full scale version of all videos. (Provide benchmarks or video_uids filters to reduce the 5TB download size.)
clipsClips available for benchmark training tasks. (Provide benchmarks or video_uids filters to reduce the download size.)
video_540ssThe downscaled version of all videos - rescaled to 540px on the short side. (Provide benchmarks or video_uids filters to reduce the 5TB download size.)
annotations_540ssThe annotations corresponding to the downscaled video_540ss videos - primarily differing only in spatial annotations (e.g. bounding boxes).
3dAnnotations for the 3D VQ benchmark.
3d_scans3D location scans for the 3D VQ benchmark.
3d_scan_keypoints3D location scan keypoints for the 3D VQ benchmark.
imuIMU data for the subset of videos available
slowfast8x8_r101_k400Precomputed action features for the Slowfast 8x8 (R101) model
omnivore_video_swinlPrecomputed action features for the Omnivore Video model
omnivore_image_swinlPrecomputed action features for the Omnivore Image model
fut_locImages and annotations for the future locomotion benchmark.
av_modelsModel checkpoints for the AV/Social benchmark.
lta_modelsModel checkpoints for the Long Term Anticipation benchmark.
moments_modelsModel checkpoints for the Moments benchmark.
nlq_modelsModel checkpoints for the NLQ benchmark.
sta_modelsModel checkpoints for the Short Term Anticipation benchmark.
vq2d_modelsModel checkpoints for the 2D VQ benchmark.


Each dataset contains a manifest.csv file that lists it's contents as well as additional metadata that's available. In particular, for full_scale there is metadata for each video available. While the top level metadata ego4d.json is generally easier to consume and contains more information, you can consume most simple metadata from the manifest itself for each dataset.


The following university IDs can be specified: