Eye Movement Recognition: Exploring Trade-Offs in Deep Learning Approaches with Development

Eye movement recognition has garnered substantial attention in recent years across diverse disciplines such as Human-Computer Interaction (HCI), medical diagnostics, and assistive technologies. This technology offers transformative possibilities, especially for individuals with paralysis and disabilities. Yet, the deployment of deep learning models for eye movement classification using non-intrusive head-free cameras like webcams remains fraught with challenges. These challenges include the lack of comparative and benchmarking studies that guide researchers and practitioners in choosing appropriate deep learning models that match such complex tasks. To address these challenges, we conducted a meticulous comparative analysis of selected deep learning architectures, including customized and fine-tuned versions of ResNet-18, EfficientNet-B0, and AlexNet. Our analysis primarily aims to evaluate the performance and generalizability for each model across a complex dataset encompassing various conditions. The n-fold cross-validation is employed to assess the robustness of our findings. Our empirical assessments reveal a nuanced landscape. For instance, ResNet-18 excels in terms of accuracy with 99.5% and acquires a competitive small model size of 43MB, while AlexNet acquires around 222MB. While this advantage comes with slightly higher computational and memory overhead compared to models like EfficientNet-B0. This study offers critical insights into the trade-offs involved in selecting an optimal deep learning model for eye movement recognition under real-world conditions.