A Mechanistic Transform Model for Synthesizing Eye Movement Data with Improved Realism

TL;DR

This paper introduces an advanced model for synthesizing more realistic degraded eye movement data by mimicking low-quality device signals, enhancing the utility of high-quality recordings for various applications.

Contribution

The paper presents a novel mechanistic transform model that improves the realism of synthetic eye movement data by degrading spatial and temporal accuracy, with a new distribution-matching technique.

Findings

Improved realism of synthetic signals demonstrated on different eye trackers.

Median classification accuracy increased by 35.7% using the new model.

Provides an application-agnostic workflow for realism assessment.

Abstract

This manuscript demonstrates an improved model-based approach for synthetic degradation of previously captured eye movement signals. Signals recorded on a high-quality eye tracking sensor are transformed such that their resulting eye tracking signal quality is similar to recordings captured on a low-quality target device. The proposed model improves the realism of the degraded signals versus prior approaches by introducing a mechanism for degrading spatial accuracy and temporal precision. Moreover, a percentile-matching technique is demonstrated for mimicking the relative distributional structure of the signal quality characteristics of the target data set. The model is demonstrated to improve realism on a per-feature and per-recording basis using data from an EyeLink 1000 eye tracker and an SMI eye tracker embedded within a virtual reality platform. The model improves the median classification accuracy performance metric by 35.7% versus the benchmark model towards the ideal metric of 50%. This paper also expands the literature by providing an application-agnostic realism assessment workflow for synthetically generated eye movement signals.