Improving Deep Learning-Based Eye Movements Classification Using Bayesian Optimization

Eye tracking technology has emerged as a touchless solution for gaze-based object selection, holding immense promise in the field of assistive technology. Despite its potential, accurately classifying various eye movements remains a formidable challenge, critical for dependable object selection. This paper introduces an innovative approach to enhance the performance of deep learning-based eye movements classification. We leveraged Bayesian Optimization (BO) to optimize the Temporal Convolutional Networks (TCNs), addressing a critical gap in prior research by optimizing hyperparameters. BO, a model-based optimization technique, efficiently explores the hyperparameter search space that leads to significant improvements in classification performance. To rigorously assess our approach, we conducted experiments on the GazeCom dataset, a rich resource annotated for diverse eye movements with a specific emphasis on smooth pursuit that is vital for calibration-free eye tracking applications. Using a lighter model, our approach significantly improved the classification of different types of eye movements—including fixation, saccade, and smooth pursuit. This result outperformed the baseline TCNs model by a margin of 1% to 7.21%. A notable improvement was observed in the classification result of smooth pursuit eye movement (F1 score: 0.8346). This achievement marks a decisive step toward refining the performance of assistive technology based on gaze interaction. Furthermore, our study can be used as a guide for future implementation of hyperparameters optimization in deep learning-based eye movements classification.