LISA: Language-guided Interference-aware Spatial-Frequency Attention for Driver Gaze Estimation

2026-05-17

AI Summary

This work addresses the degradation in driver gaze estimation accuracy caused by abrupt illumination changes, sensor noise, and irrelevant visual attributes. To this end, the authors propose LISA, a novel framework that uniquely integrates frequency-domain priors with vision-language knowledge through a disturbance-aware spatial-frequency attention mechanism. LISA fuses spatial and frequency domains to inject low-frequency semantic stability into high-frequency details while employing spatial attention to focus on eye regions. Furthermore, it leverages a frozen CLIP encoder combined with orthogonal regularization to disentangle gaze-relevant features from appearance-related distractions during training. Evaluated on two benchmark datasets, LISA demonstrates significantly enhanced robustness against occlusions and illumination variations, achieving state-of-the-art performance.

Abstract

Driver gaze estimation serves as a fundamental metric for evaluating driver attentiveness in modern monitoring systems. Beyond being vulnerable to sudden lighting changes and sensor noise, spatial-domain models struggle to disentangle authentic gaze cues from irrelevant visual attributes. In this paper, we propose LISA, a \textbf{L}anguage-guided \textbf{I}nterference-aware \textbf{S}patial-Frequency \textbf{A}ttention framework that combines frequency-domain priors with vision-language knowledge. Observing that the amplitude spectrum remains relatively stable even under spatial perturbations, we design a dual-domain fusion mechanism. It integrates stable low-frequency semantics into high-frequency details, employing spatial attention to precisely target ocular regions. To reduce semantic ambiguity, we also introduce a training-time disentanglement strategy. Using a frozen CLIP encoder and orthogonal regularization, we explicitly separate gaze features from appearance interference. Experiments on two benchmarks show that LISA achieves state-of-the-art performance, with significantly improved robustness against occlusions and lighting variations. The code repository is available at https://github.com/Mason-bupt/LISA.

Problem

• driver gaze estimation
• spatial interference
• lighting variations
• semantic ambiguity
• visual attributes

Innovation

• spatial-frequency attention
• vision-language guidance
• frequency-domain priors
• feature disentanglement
• gaze estimation