Dataset / arXiv / Oct 13, 2021
Ego4D: Around the World in 3,000 Hours of Egocentric Video
Ego4D established a large-scale benchmark for first-person perception with 3,670 hours of daily-life video from 931 wearers across 74 locations in nine countries.
Ego4D is a large-scale egocentric video dataset and benchmark suite designed around everyday first-person experience. Its breadth spans household activity, work, leisure, outdoor scenes, social interaction, and hand-object manipulation.
The paper reports 3,670 hours of video from 931 camera wearers across 74 locations in nine countries. This scale lets researchers study how actions and environments vary across people, regions, tasks, and recording conditions.
Parts of the collection include additional audio, three-dimensional scene information, gaze, stereo, and synchronized recordings from multiple egocentric cameras. These modalities belong to the Ego4D dataset and support its specific benchmark design.
The benchmark tasks organize first-person understanding across past, present, and future: episodic-memory queries, hand-object interaction and audiovisual conversation, and activity forecasting. Together they show how long-form wearable video can support multiple research questions from one collection.
The cited arXiv paper documents the collection design, benchmark definitions, consortium, annotations, and evaluation results.