Skip to main content

CLI Downloader


As a pre-requisite, you must:

  1. Make sure you have a License Agreement. If you don't, please refer to the Getting Started page. The License Agreement process takes 2 days: so do it early!
  2. Have setup or installed the Ego4D codebase, please refer to the Ego4D Github README.
  3. Configure your aws client, please see below.
  1. Upgrade your ego4d pip package to 1.7.1 or above to enable faster downloads
  2. Ensure you have setup your aws client

The dataset is large, due to this it is partitioned into separate "parts" (subsets). You may choose to download one or more parts of the dataset with optional filters in order to reduce how much you need to download.

The different parts of the dataset include:

PartSize (GB)Description
metadata0.046See metadata
annotations10.528All the annotations in Ego-Exo4D
takes10553.486Frame aligned video files associated to the takes
captures43.618Timesync and post-survey data at the capture level (multiple takes)
take_trajectory509.503Trajectories trimmed at each take
take_eye_gaze3.265Eye gaze for each take (3D & 2D)
take_point_cloud6164.615Point clouds for each take
take_vrs12301.458VRS files for each take
take_vrs_noimagestream995.592VRS files for each take without image stream data (video data within MP4 containers with --parts takes)
capture_trajectory851.691Trajectory at the capture-level
capture_eye_gaze5.619Eye gaze at the capture-level (3D)
capture_point_cloud4750.039Point clouds for each capture
downscaled_takes/448438.556Downscaled takes at 448px on the shortest side
features/omnivore_video49.986Omnivore video features
features/maws_clip_2b533.826MAWS CLIP (ViT-2b) features for each frame of video
ego_pose_pseudo_gt138.629Pseudo-ground truth data for Ego Pose
expert_commentary42.292Commentaries for each expert (audio recordings)
take_transcription0.094Audio transcriptions for each take
take_audio1056.907Audio files for the egocentric aria camera
all38449.748All data within the release (you can use --parts all)
default12112.773The default set of data in the release (you can use --parts default or provide no parts)

The parts in bold refer to the recommended set (default). You can additionally filter based on:

  • Data relevant to a specific benchmarks: --benchmarks <benchmark_name>
    • e.g. --benchmarks keystep relations
  • Dataset splits (train/val/test): --splits <s1> <s2> ..., e.g.
    • --splits train for data that only lies in the training set
  • Egocentric or exocentric related data (videos, trajectory)
    • Use --views exo for only exocentric, --views ego for only egocentric, or --views ego exo for both. Defaults to both.
  • Specific takes or captures to download: --uids
  • Universities (use --universities or -u)
    • e.g. -u cmu unc sfu

Please refer to the Download CLI README or use the flag --help for the latest parts of the dataset and additional usage of the dataset.

Setup AWS Client

To setup your aws client, you will need to:

  1. Install the aws-cli pip package, i.e. pip install awscli
  2. Run aws configure or (see the official docs)

When running aws configure you simply need to enter your access key & secret key and then press Enter twice to follow the default region and output format.

You can have multiple profiles setup on your machine. To tell the downloader which profile to use, simply use the flag --s3_profile <name>.

Download The Data


Before you download data, please setup your aws client.

Below are some examples on how to run the downloader CLI with common options:

As noted above, you can use --help to view all the up-to-date options
egoexo --help

Download suggested dataset parts for all the takes present.

egoexo -o <out-dir>

This is equivalent to

egoexo -o <out-dir> --parts default

Other Examples

Download all annotations & metadata:

egoexo -o <out-dir> --parts annotations metadata

Download all point clouds, eye gaze and trajectories for each take:

egoexo -o <out-dir> --parts take_point_cloud take_eye_gaze trajectory

Download all point clouds, eye gaze and trajectories:

egoexo -o <out-dir> --parts take_point_cloud eye_gaze take_trajectory