Meta Builds Massive First-Person Video Dataset to Train AI Vision
Meta unveils Ego4D, a massive dataset of first-person videos designed to teach artificial intelligence to see the world from a human perspective.
Meta introduces Ego4D, a massive new dataset that uses first-person video to teach artificial intelligence how to perceive the world as humans do. By partnering with 13 universities across nine countries, the company collects 2,200 hours of footage from 700 participants going about their daily lives.
Unlike traditional computer vision projects that rely on a bystander's perspective, this footage captures point-of-view experiences like social interactions, hand manipulation, and activity prediction. Lead research scientist Kristen Grauman states that this stands as the largest dataset ever created specifically for this first-person focus.
The company plans to release the data to researchers in November and will host a challenge next year to push this technology forward. This effort naturally aligns with Meta's recent Ray-Ban smart glasses partnership, though it also raises familiar privacy concerns regarding the potential for increased surveillance.