A Study on the Impact of Different Components of a Traditional Webcam-Based 2-D Gaze-Tracking Algorithm

Webcam-based 2D gaze tracking algorithms are lightweight and are becoming increasingly used in the fields of medicine, market research and many others. As they become increasingly used, it becomes vital to break down their components to understand their limitations and better explore their practical implications. Key components of the gaze tracking pipeline are the calibration pattern, landmark detector, eye patch generation method, and the final eye-gaze model. Through an experimental framework, this work explores various methods for these components and evaluates the impact of each component on the final performance of an individualised real-time gaze tracking algorithm that is trained and tested on data from single individuals, as opposed to generalised approaches that are trained on data from multiple individuals. Gaze tracking data from users looking at a laptop screen were captured using a webcam and were used for the evaluation of the examined methods. The final proposed pipeline for individualised webcam-based real-time gaze-tracking under “real-world” use cases achieved a 2.26 cm accuracy compared to 3.42 cm for similar approaches. Additional validation on an independent publicly available dataset (EyeDIAP) further supports our findings.