Discount Eye-Tracking for Usability Testing with Webcam, CNN, and the Selected Training Data

Not many people realize that eye-tracking might be a fruitful addition to the widely established usability testing, at the same time not becoming prohibitively expensive for smaller companies. Unlike in research applications of the human gaze training technologies that require high precision, usability engineering professionals are mostly interested in concrete improvements to the user interfaces they develop. Correspondingly, affordable budgets and more frequent testing sessions yield better practical results. This paper presents a “discount” approach implementing “good enough” eye-tracking for usability testing with a standard webcam. A convolutional neural network was trained to predict screen coordinates from eye images, compensating for the precision provided by the specialized equipment. To improve the accuracy, three different data collection methods were tested, and the combined mesh-point based one allowed the best error (Root Mean Square Error, RMSE) of 78 pixels on a 1920xl080 screen. Although the error was about 2–3 inferior to the one obtained for the professional equipment, our results demonstrate that a webcam combined with deep learning techniques can serve as an affordable alternative for real-time eye-tracking. This approach has the potential to significantly reduce costs and make eye-tracking technology more accessible to a broader community of software developers and user interface designers.