A Novel Eye Center Localization Method for Head Poses With Large Rotations

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Accurate eye localization is vital for human-computer interaction and information acquisition.
• Existing eye localization methods struggle with large head rotations and occlusions, limiting their real-world applicability.

Purpose of the Study:

• To develop novel methods for accurate eye center localization, addressing limitations of existing approaches in challenging conditions.
• To introduce a complete representation (CR) pipeline for enhanced eye localization.

Main Methods:

• Proposed a complete representation (CR) pipeline generating CR-center and CR-region representations.
• Developed two eye center localization methods: one using geometric transformation and CR-center, the other using image translation and CR-region with generative adversarial networks.

Main Results:

• The proposed methods demonstrated superior accuracy and robustness in localizing eye centers compared to existing techniques.
• Successfully localized eye centers in images with significant head rotation (yaw: -67.5° to +67.5°, roll: +120° to -120°), occlusions, and poor illumination.

Conclusions:

• The novel CR pipeline and associated eye localization methods significantly improve performance in challenging scenarios.
• This research advances the capability of machines to understand human gaze and interaction in diverse and difficult conditions.