Gaze prediction forecasts a viewer's near-future gaze position from an eye-tracking signal so systems such as foveated rendering can pre-render where the eye will land. This concept examines how prediction accuracy depends on the type of eye movement (fixations, small and large saccades, and the post-saccade critical evaluation period) and on the individual viewer, comparing an LSTM, a transformer, and an oculomotor-plant Kalman-filter model. The central lesson is that mean accuracy hides large subject-to-subject variation, and that oculomotor measures such as fixation-velocity noise and saccade velocity predict which viewers are hardest to model.
This study compares three models for predicting a viewer's near-future gaze position from a 1000 Hz eye-tracking signal: a lightweight LSTM, a transformer-based time-series model (TST), and an Oculom…