On current limitations of online eye-tracking to study the visual processing of source code

Abstract

Context:

Eye-tracking is an increasingly popular instrument to study how programmers process and comprehend source code. While most studies are conducted in controlled environments with lab-grade hardware, it would be desirable to simplify and scale participation in experiments for users sitting remotely, leveraging home equipment.

Objective:

This study investigates the possibility of performing eye-tracking studies remotely using open-source algorithms and consumer-grade webcams. It establishes the technology’s current limitations and evaluates the quality of the data collected by it. We conclude by recommending ways forward to address the shortcomings and make remote code-reading studies in support of eye-tracking feasible in the future.

Method:

We gathered eye-gaze data remotely from 40 participants performing a code reading experiment on a purpose-built web application. The utilized eye-tracker worked client-side and used ridge regression to generate x- and y-coordinates in real-time predicting the participants’ on-screen gaze points without the need to collect and save video footage. We processed and analysed the collected data according to common practices for isolating eye-movement events and deriving metrics used in software engineering eye-tracking studies. In response to the lack of an algorithm explicitly developed for detecting oculomotor fixation events in low-frequency webcam data, we also introduced a dispersion threshold algorithm for that purpose. The quality of the collected data was subsequently assessed to determine the adequacy and validity of the methodology for eye-tracking.

Results:

The collected data was found to be of varying quality despite extensive calibration and graphical user guidance. We present our results highlighting both the negative and positive observations from which the community hopefully can learn. Both accuracy and precision were low and ultimately deemed insufficient for drawing valid conclusions in a high-precision empirical study. We nonetheless contribute to identifying critical limitations to be addressed in future research. Apart from the overall challenge of vastly diverse equipment, setup, and configuration, we found two main problems with the current webcam eye-tracking technology. The first was the absence of a validated algorithm to isolate fixations in low-frequency data, compromising the assurance of the accuracy of the data derived from it. The second problem was the lack of algorithmic support for head movements when predicting gaze location. Unsupervised participants do not always keep their heads still, even if instructed to do so. Consequently, we frequently observed spatial shifts that corrupted many collected datasets. Three encouraging observations resulted from the study. Even when shifted, gaze points were consistently dispersed in patterns resembling both the shape and size of the stimuli without extreme deviations. We could also distinguish recognizable reading patterns. Linearity was significantly different when participants were reading source code compared to natural text, and we could detect the expected left-to-right and top-to-bottom reading directions for participants reading natural text snippets.

Conclusion:

The accuracy and precision levels were not sufficient for a word-by-word analysis of code reading but could be adequate for a broader, coarse-grained precision study. Additionally we identified two main issues compromising the collected data validity and contributed a fixation detection algorithm to approach one of these issues. With suitable solutions to the identified issues, remote eye-tracking studies with webcams on code reading could eventually be feasible.