Facebook Builds Massive First-Person Video Dataset to Train Future AI
Facebook announces Ego4D, a massive 2,200-hour first-person video dataset designed to train artificial intelligence for smart glasses and virtual reality. The tech giant plans to release the data publicly to researchers in November.
Facebook announces a new research project called Ego4D, which consists of 2,200 hours of first-person video footage collected from around the world. This massive dataset aims to train next-generation artificial intelligence models by showing them exactly what a human sees in everyday life. With 20 times more footage than any similar existing dataset, Ego4D represents a major leap forward in AI training resources.
This first-person footage directly supports Facebook's Reality Labs division and its heavy investments in wearable technology. The data teaches AI to understand and identify objects from a human perspective, which enhances the functionality of devices like the recently released Ray-Ban Stories smart glasses and Oculus VR headsets. Furthermore, researchers suggest this video data helps robots learn about the physical world without requiring manual, step-by-step programming.
Facebook collaborates with 13 university partners across nine countries to gather this footage from over 700 participants. In November, the company makes the Ego4D dataset publicly available to the global research community. Lead scientists believe this open-access approach sparks rapid innovation, allowing academics and external developers to solve complex AI problems at a much greater scale.