Optimizing Gaze Estimation: A Comparative Study of Loss Functions and Deep Architectures

Eye tracking, particularly gaze estimation, is becoming increasingly important in HCI, healthcare, and VR applications, as it provides deeper insights into visual behavior. In this paper, we conduct a comprehensive ablation study to analyze the architectural changes that affect gaze estimation performance. Furthermore, we experiment with different loss functions to determine if using a combined loss significantly changes the model performance. Our findings reveal that since the combined loss function offers only marginal improvement over using a single loss function alone (0.02 degrees), a single loss function may suffice in specific contexts. Additionally, transformer-based models demonstrate superior performance, offering smoother optimization and better results under challenging conditions. These insights highlight the potential for optimizing gaze estimation models by balancing loss function complexity and leveraging modern transformer architectures.