Eye Gaze Region Estimation Via Multi-Scale Sparse Dictionary Learning

Traditional model-based gaze estimation methods relying on pupil or iris detection is vulnerable under large head movement. To solve this problem, a novel head pose-free appearance-based eye gaze prediction method without gaze calibration is proposed, combining head pose information and multi-scale eye image sparse features to determine the gaze region. Multi-scale dictionaries are learned by sparse coding to represent eye appearance more effectively, in which whole eye images are employed as global scale while their corresponding distinct patch-based images as local scale. Supported by multiscale dictionaries, multi-scale eye image sparse features are the sparse coefficients encoding eye image by linear combinations of bases selected from the learned dictionaries. In order to handle gaze variation under large head movement, a multiclass classifier concatenates the known head pose information and multi-scale eye image sparse features to estimate the gaze region. The experimental results show that the proposed approach outperforms baseline methods on a public dataset.