CLI Downloader

info

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.

tip

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:

Part	Size (GB)	Description
metadata	0.046	See metadata
annotations	10.533	All the annotations in Ego-Exo4D
takes	10553.486	Frame aligned video files associated to the takes
captures	43.618	Timesync and post-survey data at the capture level (multiple takes)
take_trajectory	509.503	Trajectories trimmed at each take
take_eye_gaze	3.265	Eye gaze for each take (3D & 2D)
take_point_cloud	6164.615	Point clouds for each take
take_vrs	12301.458	VRS files for each take
take_vrs_noimagestream	995.592	VRS files for each take without image stream data (video data within MP4 containers with --parts takes)
capture_trajectory	851.691	Trajectory at the capture-level
capture_eye_gaze	5.619	Eye gaze at the capture-level (3D)
capture_point_cloud	4750.039	Point clouds for each capture
downscaled_takes/448	438.556	Downscaled takes at 448px on the shortest side
features/omnivore_video	49.986	Omnivore video features
features/maws_clip_2b	533.826	MAWS CLIP (ViT-2b) features for each frame of video
ego_pose_pseudo_gt	138.629	Pseudo-ground truth data for Ego Pose
expert_commentary	42.292	Commentaries for each expert (audio recordings)
take_transcription	0.094	Audio transcriptions for each take
take_audio	1056.907	Audio files for the egocentric aria camera
all	38449.753	All data within the release (you can use --parts all)
default	12112.778	The 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 taskgraph
• 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

info

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

Recommended Dataset

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