Testing multiple polynomial models for eye-tracker calibration

Abstract

The straightforward approach to eye-tracker calibration considers that the calibration data do not have erroneous associations, and the calibration function is defined. The violation of the non-erroneous assumption could cause an arbitrarily large bias. The MMransac algorithm proposed in this paper is a modified version of the Random Sample Consensus. that achieves robust calibrations. On the other hand, polynomials in two variables (i.e., with terms in the form κxayb) are commonly used to map eye-tracker measurements to points on the screen. High-degree polynomials tend to be more accurate; however, as the order is increased, the function becomes more complex and less smooth, which could cause over-fitting. In this sense, this paper proposes an algorithmic approach that enables model selection criteria even in the presence of outliers. This approach was tested using different model selection criteria. Results show that more accurate calibrations are obtained with the combined robust fitting and model selection approach using the Akaike information criterion (AIC) and the Kullback information criterion (KIC).