Person-independent 3D Gaze Estimation using Face Frontalization

Abstract

Person-independent and pose-invariant estimation of eye-gaze is important for situation analysis and for automated video annotation. We propose a fast cascade regression based method that first estimates the location of a dense set of markers and their visibility, then reconstructs face shape by fitting a part-based 3D model. Next, the reconstructed 3D shape is used to estimate a canonical view of the eyes for 3D gaze estimation. The model operates in a feature space that naturally encodes local ordinal properties of pixel intensities leading to photometric invariant estimation of gaze. To evaluate the algorithm in comparison with alternative approaches, three publicly-available databases were used, Boston University Head Tracking, Multi-View Gaze and CAVE Gaze datasets. Precision for head pose and gaze averaged 4 degrees or less for pitch, yaw, and roll. The algorithm outperformed alternative methods in both datasets.