USGaze: Temporal Gaze Estimation via a Unified State-Space Modeling Framework

Existing appearance-based and video-based gaze estimation methods mainly rely on frame-wise prediction or local-window temporal fusion, which limits their ability to model long-range dependencies and to explicitly suppress output-level jitter. This leaves a gap in unified temporal gaze estimation frameworks that jointly address contextual feature aggregation and prediction-level stabilization. To address this limitation, we propose a unified state-space temporal gaze estimation framework to improve both angular accuracy and temporal consistency. Specifically, consecutive eye image sequences are mapped into a shared latent state space, where spatial appearance cues and inter-frame dynamics are jointly modeled. A feature-level temporal aggregation module is further designed to adaptively reweight historical observations for the current estimate, and a prediction-level temporal correction module is introduced to suppress short-term fluctuations while preserving rapid gaze shifts. On the TEyeD dataset after quality screening, the proposed method achieves a 3D gaze MAE of 0.533°, compared with 0.96° for Model-aware and 3.18°–3.47° for the ResNet baselines reported in the original TEyeD paper, while maintaining manageable deployment overhead. These results indicate that the proposed framework provides a favorable balance between estimation accuracy, temporal stability, and practical efficiency.