Improving the Accuracy of Low Quality Eye Trackers

Abstract

The tracking of the human eye-gaze has been shown to have several potential applications. For instance, it has been shown that in the elderly, the patterns with which the eye recognises faces can be used as a predictor of cognitive decline with good accuracy. Good quality eye-tracking hardware, which is currently used for eye-tracking, is often expensive and bulky/invasive. Increasing the accessibility of eye-tracking hardware is therefore important so that it can be used more widely as a diagnostic tool and in other commercial applications e.g in human computer interaction, market research analysis for webpages etc. Work has been done in recent years to make this possible by making eye-trackers out of commodity hardware such as laptop webcams and tablet cameras. These approaches use software-based methods to predict the corresponding measurement of a high-quality eye-tracker given a measurement from the commodity hardware. Our aim in this project is to improve these of these models so that the accuracy of the commodity hardware-based eye tracker can be improved further. Selected earlier models used in this software-based improvement approach include a convolutional neural network-based model and a Kalman smoothing probabilistic time-series model. Using the iTracker model developed by Krafka, Khosla et. al (2016), and further developed by You (2018) as a benchmark, we proceed to try new features and ways of training a neural network to handle the prediction task. Using a simple encoder architecture, we manage to process the results from the iTracker that had been ported to the desktop and achieve performance improvements over the chosen baseline.