Facebook Builds Massive First-Person Video Dataset to Train Future AI
Facebook collaborates with global universities to release Ego4D, a dataset of 3,000 hours of first-person video designed to teach AI how to understand the world from a human perspective.
Facebook releases a massive new dataset called Ego4D to help artificial intelligence systems understand the world from a first-person perspective. Current AI models struggle to recognize everyday activities when viewed through the eyes of a person, so Facebook teams up with 13 partner universities to gather thousands of hours of video from over 700 participants across nine countries.
The research partners capture footage of common activities like cooking, grocery shopping, and tying shoelaces using glasses cameras and GoPros. A dedicated research team carefully edits and hand-annotates the footage, narrowing the collection down to 3,000 hours of high-quality video that also includes data from staged environments.
Facebook makes the Ego4D dataset freely available to the global research community to accelerate the development of AI for augmented reality glasses and other wearables. Researchers hope this first-person approach teaches AI to understand daily life activities through human eyes, marking a significant shift away from traditional third-person computer vision training.