Gaze Estimation and Interaction in Real-World Environments

fast, pervasive, and non-verbal interaction between humans and computers. Appearance-based gaze estimation
methods directly regress from eye images to gaze targets without eye feature detection, and therefore have great
potential to work with ordinary devices. However, these methods require a large amount of domain-specific
training data to cover the significant variability in eye appearance. In this thesis, we focus on developing
appearance-based gaze estimation methods and corresponding attentive user interfaces with a single webcam
for challenging real-world environments. We collected a large-scale gaze estimation dataset from real-world
setting, proposed a full-face gaze estimation method, and studied data normalization. We applied our gaze
estimation methods to real-world interactive application including eye contact detection and gaze estimation
with multiple personal devices.