Hover your mouse over scene cameras above to see a sample video for the chosen scenario.
A diverse, large-scale multi-modal, multi-view, video dataset and benchmark collected across 13 cities worldwide by 740 camera wearers, capturing 1286.3 hours of video of skilled human activities.
5035
(# takes)
740
(# participants)
123
(# sites)
1286.3
(# ego+exo hours)
Cooking
Music
Soccer
Health
Basketball
Dance
Bike Repair
Rock Climbing
For further detailed information refer to our manuscript:
K Grauman et al. Ego-Exo4D: Understanding Skilled Human
Activity from First- and Third-Person Perspectives.
PDF
|
Supplementary PDF
|
ArXiv paper
The dataset is now publicly available. Please refer to the Getting Started page on the Documentation Site for details on how to download the dataset. To summarize you will need to:
Yes. As new university partners are part of this effort, a new license should be signed.
The data was collected from 740 participants. We showcase a distribution of age and gender from our participants who volunteered to self-identify their demographics — age and gender.
Yes. Since the scenarios allow for closed environments (e.g., no passerby) nearly all video is available without de-identification. Refer to our paper for details of our privacy and data collection pipeline.
A sample visualisation of our scenarios is below.
We will be releasing additional information with the dataset before the end of the year
Ego4D collected unscripted daily life activities. Instead, we here focus on skilled activities specifically, ensuring participants of diverse levels of skill are recorded performing the same scenario, while maintaining distinctions natural to the scenario being carried out in a new environment. We cover both procedural and physical skilled activities. While some of these general scenarios have been captured in Ego4D (there are indeed examples of cooking, climbing or soccer), these are disjoint and do not record the same scenario making them unsuitable for tasks related to to keystep recognition or skill assessment.
Importantly, Ego-Exo4D is a multi-view dataset with synchronized egocentric and multiple exocentric view. This is the largest and most diverse effort for multi-view ego-exo data.
The egocentric views are captured using the Aria Project devices.
Yes! Please refer to challenges page on the docs.
Minneapolis, U.S.
Bloomington, U.S.
U.S.
U.S.
U.S.
Canada
International
FAIR
Aria
HALO
Research Intern
The dataset was released on 13th December 2023 and is now publicly available
License forms should be signed to access & download the dataset.
Sign Ego-Exo4D Licenses ↗
Refer to "Getting Started" on the Documentation Site
Getting Started ↗
