Gaze Estimation by Exploring Two-Eye Asymmetry.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Accurate eye gaze estimation is crucial for intelligent systems and interactive applications.
• Estimating gaze direction from a single eye image presents significant challenges due to complex regression.
• Existing methods often struggle with the inherent asymmetry between a person's two eyes.

Purpose of the Study:

• To develop an optimized eye gaze estimation method that accounts for two-eye asymmetry.
• To improve the efficiency and accuracy of gaze direction prediction in intelligent systems.
• To introduce a novel network architecture that leverages differential eye performance.

Main Methods:

• Proposed the Face-based Asymmetric Regression-Evaluation Network (FARE-Net), comprising a Face-based Asymmetric Regression Network (FAR-Net) and an Evaluation Network (E-Net).
• FAR-Net predicts 3D gaze directions using an asymmetric mechanism that weights and sums losses from both eyes.
• E-Net learns eye reliabilities to balance asymmetric and symmetric learning mechanisms.

Main Results:

• FARE-Net achieved state-of-the-art performance on benchmark datasets including MPIIGaze, EyeDiap, and RT-Gene.
• The asymmetric mechanism effectively utilized high-performing eyes to optimize network training.
• Error distribution analysis and ablation studies confirmed the effectiveness of the proposed method.

Conclusions:

• The proposed FARE-Net effectively addresses the challenge of two-eye asymmetry in gaze estimation.
• This approach leads to significant improvements in gaze estimation accuracy and reliability.
• FARE-Net offers a promising solution for advanced interactive intelligent systems.