Gaze Estimation From Eye Appearance: A Head Pose-Free Method via Eye Image Synthesis

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Machine Learning

Background:

• Appearance-based gaze estimation is hindered by significant changes in eye appearance due to free head motion.
• Conventional methods struggle as training images from one head pose are ineffective for testing at other poses.

Purpose of the Study:

• To develop a novel gaze estimation method capable of handling free head motion using eye image synthesis from a single camera.
• To improve the robustness and accuracy of gaze estimation systems in real-world scenarios with unconstrained head movements.

Main Methods:

• Propose a single-directional (SD) flow model to manage eye image variations caused by head motion.
• Synthesize new training images for unseen head poses using reference head poses and estimated SD flows.
• Employ joint optimization for simultaneous eye image alignment and gaze estimation using synthetic data.

Main Results:

• The proposed method effectively synthesizes training images for diverse head poses, overcoming limitations of fixed-pose training.
• Experimental evaluations demonstrate the method's effectiveness and performance in handling free head motion for gaze estimation.
• The SD flow model efficiently addresses eye appearance variations, contributing to improved accuracy.

Conclusions:

• The novel eye image synthesis approach enables robust appearance-based gaze estimation under free head motion.
• This method offers a significant advancement for single-camera gaze tracking systems by adapting to unconstrained head movements.
• The proposed technique provides a practical solution for real-time gaze estimation applications.