Temporal smoothness and broader continuous-stream generalization of GazeFS

Characterize the temporal smoothness and broader continuous-stream generalization of GazeFS beyond the segmented, successful target-acquisition episodes and evaluation conditions reported in the study.

Background

GazeFS is an online model that forecasts the next target-center direction and a short-horizon Search/Focus phase from gaze–head history without using target geometry at inference. The reported experiments demonstrate sustained-Focus reductions in target bias, dispersion, median error, and upper-tail error, but they also show a Raw-relative motion cost, indicating that spatial centering and temporal smoothness are not simultaneously optimized by the current system.

The evaluation is based primarily on segmented successful acquisitions from a single HoloLens 2 task, with a secondary closed-loop study. The paper therefore leaves unresolved whether GazeFS can provide sufficiently smooth temporal behavior and whether its benefits generalize to broader continuous-stream interaction settings beyond the recorded task and device conditions.