Gaze Estimation Under Low-Resolution Conditions: A Hybrid Training Framework with Frequency-Aware Alignment and Adaptive Distillation

Appearance-based gaze estimation in far-field, device-free Human–Computer Interaction settings (e.g., IDEs) is often impaired by low effective resolution, where gaze-discriminative ocular details are lost, and predictions become unstable. Existing low-resolution treatments frequently apply indiscriminate feature alignment or uniform distillation, which may transfer low-frequency nuisance cues (e.g., illumination) and suffer from negative transfer when teacher predictions are unreliable. We propose a teacher–student training framework for robust 3D gaze estimation under low-resolution inputs, featuring (1) frequency-aware selective distillation that emphasizes mid-to-high frequency structures, (2) uncertainty-weighted knowledge transfer that down-weights low-confidence supervision, and (3) an adaptive multi-resolution curriculum with resolution-dependent distillation strength. We implement the method and evaluate it on Gaze360 and RT-GENE under a controlled low-resolution protocol (×2/ × 4/ × 8), reporting Mean Angular Error (MAE) across degradation levels. Improved robustness is expected to enhance interaction stability by reducing gaze-driven mis-triggers and the need for repeated confirmation or manual recovery, while extremely severe degradations that eliminate discriminative structures remain challenging and motivate future robustness and interaction-level studies.