Eye Tracking on Your Cellphone (gazecapture.csail.mit.edu)

A groundbreaking advancement in eye tracking technology has been announced with the introduction of GazeCapture, a comprehensive dataset designed to make eye tracking widely accessible. This dataset consists of data from over 1,450 individuals and approximately 2.5 million frames, enabling the development of iTracker—a convolutional neural network specifically tailored for real-time eye tracking on standard mobile devices. This innovation signifies a major step forward, as it eliminates the need for additional sensors or specialized hardware, allowing consumers to leverage eye tracking capabilities on their existing smartphones and tablets. The significance of this development lies in iTracker's impressive performance metrics: it achieves a prediction error of only 1.7 cm on mobile phones and 2.5 cm on tablets without the need for calibration, and even lower errors of 1.3 cm and 2.1 cm with calibration, all while maintaining real-time processing speeds of 10-15 frames per second. Importantly, the model's learned features demonstrate strong generalization capabilities across different datasets, marking a new era in machine learning applications for eye tracking. This research, supported by major tech companies, opens up potential for advancements in various fields, including user experience design, accessibility, and behavioral research.