One of the primary objectives of computer vision projects is to enable machines to perceive the world like humans do. Facebook is addressing this challenge with its new initiative, Ego4D, which involves a comprehensive dataset designed to train future AI models.
Facebook has partnered with 13 universities across nine countries to gather 2,200 hours of footage from 700 participants. This footage captures experiences from the user’s perspective, making it invaluable for developing AI systems that understand human behavior. According to Kristen Grauman, Facebook’s lead research scientist, this collection represents the largest dataset specifically focused on this goal.
The footage highlights a variety of everyday human experiences, such as social interactions, hand movements, and object manipulation, as well as predicting future events. This marks a significant shift for the social network, which has typically relied on data collected from an observer's perspective.
In November, Facebook plans to release these datasets to researchers who agree to the Ego4D data use agreement. Next year, broader research communities will be invited to further train machines to comprehend human activities and interactions.
While this initiative has the potential to enhance machine learning capabilities, it raises concerns about privacy and surveillance, especially given Facebook's partnership with Ray-Ban on camera glasses. These developments underscore the need for careful consideration of the implications of such technology in our daily lives.