Abstract
This study evaluates a smartphone-based, deep-learning eye-tracking algorithmby comparing its performance against a commercial infrared-based eye tracker,the Tobii Pro Nano. The aim is to investigate the feasibility ofappearance-based gaze estimation under realistic mobile usage conditions. Keysensitivity factors, including age, gender, vision correction, lightingconditions, device type, and head position, were systematically analysed. Theappearance-based algorithm integrates a lightweight convolutional neuralnetwork (MobileNet-V3) with a recurrent structure (Long Short-Term Memory) topredict gaze coordinates from grayscale facial images. Gaze data were collectedfrom 51 participants using dynamic visual stimuli, and accuracy was measuredusing Euclidean distance. The deep learning model produced a mean error of17.76 mm, compared to 16.53 mm for the Tobii Pro Nano. While overall accuracydifferences were small, the deep learning-based method was more sensitive tofactors such as lighting, vision correction, and age, with higher failure ratesobserved under low-light conditions among participants using glasses and inolder age groups. Device-specific and positional factors also influencedtracking performance. These results highlight the potential of appearance-basedapproaches for mobile eye tracking and offer a reference framework forevaluating gaze estimation systems across varied usage conditions.