Algorithms for the automated correction of vertical drift in eye-tracking data

Advancing Dynamic-Time Warp Techniques for Correcting Eye Tracking Data in Reading Source Code

Background: Automated eye tracking data correction algorithms such as Dynamic-Time Warp always made a trade-off between the ability to handle regressions (jumps back) and distortions (fixation drift). At the same time, eye movement in code reading is characterized by non-linearity and regressions. Objective: In this paper, we present a family of hybrid algorithms that aim to handle both regressions and distortions with high accuracy. Method: Through simulations with synthetic data, we replicate known eye movement phenomena to assess our algorithms against Warp algorithm as a baseline. Furthermore, we utilize two real datasets to evaluate the algorithms in correcting data from reading source code and see if the proposed algorithms generalize to correcting data from reading natural language text. Results: Our results demonstrate that most proposed algorithms match or outperform baseline Warp in correcting both synthetic and real data. Also, we show the prevalence of regressions in reading source code. Conclusion: Our results highlight our hybrid algorithms as an improvement to Dynamic-Time Warp in handling regressions.

Best practices for cleaning eye movement data in reading research

One challenge that comes with studying eye movement behavior is deciding how to clean the eye movement data (e.g., fixation durations) before conducting analyses. Reading researchers must decide which data cleaning methods they will use and which thresholds they will set to remove eye movements that are not reflective of lexical processing. The purpose of this project was to determine what data cleaning methods are typically used and if there are any consequences of using different data cleaning methods. In the first study, an analysis of 192 recently published articles indicated that there is inconsistency in the reporting and application of data cleaning methods. In the second study, three different data cleaning methods were applied based on the literature analysis in the first study. Analyses were conducted to determine the impact of different data cleaning methods on three commonly studied effects in reading research (frequency, predictability, and length). Overall, standardized estimates decreased for each effect when more data were removed; however, removing more data also resulted in decreased variance. As a result, effects remained significant with each data cleaning method, and simulated power remained high for both a moderate and small sample size. Effect sizes remained consistent for most effects but decreased for the length effect as more data were removed. Seven suggestions are provided that are based on open science practices with the intention of helping researchers, reviewers, and the field as a whole.

Offline Calibration for Infant Gaze and Head Tracking across a Wide Horizontal Visual Field

Most well-established eye-tracking research paradigms adopt remote systems, which typically feature regular flat screens of limited width. Limitations of current eye-tracking methods over a wide area include calibration, the significant loss of data due to head movements, and the reduction of data quality over the course of an experimental session. Here, we introduced a novel method of tracking gaze and head movements that combines the possibility of investigating a wide field of view and an offline calibration procedure to enhance the accuracy of measurements. A 4-camera Smart Eye Pro system was adapted for infant research to detect gaze movements across 126° of the horizontal meridian. To accurately track this visual area, an online system calibration was combined with a new offline gaze calibration procedure. Results revealed that the proposed system successfully tracked infants' head and gaze beyond the average screen size. The implementation of an offline calibration procedure improved the validity and spatial accuracy of measures by correcting a systematic top-right error (1.38° mean horizontal error and 1.46° mean vertical error). This approach could be critical for deriving accurate physiological measures from the eye and represents a substantial methodological advance for tracking looking behaviour across both central and peripheral regions. The offline calibration is particularly useful for work with developing populations, such as infants, and for people who may have difficulties in following instructions.