Enabling Eye Tracking for Crowd-Sourced Data Collection With Project Aria

Through a novel, multi-sensor platform in a wearable glasses form factor, paired with crowd-sourced data collection, we have enabled the collection of tens of thousands of egocentric gaze recordings across hundreds of daily tasks without the limitations posed by traditional eye-tracking or lab research. The eye tracking model that we describe and open source here was created with a large study of 1500 participants performing eye tracking calibration. Data collection utilized the device sensor suite, allowing us to leverage scene and contextual information across various environments and tasks. This paper will discuss the systems in place and technical steps we have taken to enable eye tracking for the project, and select lessons learned in the process. This large-scale data collection tool and open source eye tracking model will enable advances in mobile and pervasive eye tracking to understand nuances in eye movements and various future AI and eye-tracking applications.