This study explores the use of transfer learning in gaze estimation, focusing on the development of personalized models tailored to individual users. Our approach involves collecting gaze data using standard laptop webcams, designed to operate effectively within resource-limited settings, thereby enhancing accessibility and affordability. We conducted a comparative analysis of models using transfer learning against models trained without pre-trained weights, examining their convergence behavior and sensitivity to different dataset sizes. The analysis includes both eye and face images, providing a comprehensive view of model adaptability. Our findings show that while both methods produce comparable results overall, transfer learning offers notable advantages—particularly faster convergence, reduced computational cost, and enhanced stability when data are limited. However, the results also reveal that transfer learning is not universally superior; for face images, models trained from scratch achieved lower mean errors but exhibited higher variability, whereas transfer learning ensured more consistent performance. These insights highlight that the benefits of transfer learning depend on the data characteristics and task complexity. In the most data-constrained setting (100 images), transfer learning reduced the mean error by 20.99 px for the Left Eye model and 35.56 px for the Right Eye model, whereas for face images the models trained from scratch consistently achieved lower mean error across all evaluated dataset sizes. Overall, this study underscores the potential of transfer learning for efficient and scalable gaze estimation, particularly in small-data environments, while providing practical guidance on when and how transfer learning yields the greatest benefit for real-time applications such as human–computer interaction (HCI), assistive technologies, and personalized user experiences.
Ansari et al. (2026) studied this question.