Unsupervised Gaze Representation Learning with Conjugate Gaze Consistency Loss for Enhanced Gaze Estimation

Abstract

Patients suffering from quadriplegia and other paralysis that interfere with the ability to communicate have been increasing for the past decade, and in turn, the need for better communication boards have been growing. Current communication boards, both physical and digital communication boards, all have flaws starting from the need of an assistant to the sole expense of the device. However, gaze estimation techniques have also been gaining attention to enhance the quality of communication boards by tracking the movement of the eye via a camera to assess what the user is trying to communicate. Previous studies on gaze estimation algorithms have shown that collecting data for accurate gaze values is an arduous task, and that the accuracy of the gaze estimation models has been unsatisfactory for practical use. Thus, I propose a gaze estimation-based digital communication board system that combines gaze representation learning with transfer learning. In the representation learning phase, I introduce a random sign-reversal module to efficiently isolate gaze-related features. In the transfer learning phase, I implement a medically driven loss function to enhance the system's accuracy. The proposed system achieved an angular error of 9.42 degrees which represents state-of-the-art performance compared to previous studies.