Gazing into the Abyss: Real-time Gaze Estimation

TL;DR

This paper develops real-time gaze estimation algorithms that are accurate, computationally efficient, and require no training data, suitable for low-quality webcams on common devices.

Contribution

It introduces three gaze tracking pipelines that match modern trackers' performance without training data, bridging the gap between accuracy and real-time feasibility.

Findings

Achieved real-time gaze estimation with no training data

Matched performance of modern gaze trackers

Optimized for low-quality webcams and limited computational resources

Abstract

Gaze and face tracking algorithms have traditionally battled a compromise between computational complexity and accuracy; the most accurate neural net algorithms cannot be implemented in real time, but less complex real-time algorithms suffer from higher error. This project seeks to better bridge that gap by improving on real-time eye and facial recognition algorithms in order to develop accurate, real-time gaze estimation with an emphasis on minimizing training data and computational complexity. Our goal is to use eye and facial recognition techniques to enable users to perform limited tasks based on gaze and facial input using only a standard, low-quality web cam found in most modern laptops and smart phones and the limited computational power and training data typical of those scenarios. We therefore identified seven promising, fundamentally different algorithms based on different user features and developed one outstanding, one workable, and one honorable mention gaze tracking pipelines that match the performance of modern gaze trackers while using no training data.