Facial Landmarks Based Region-Level Data Augmentation for Gaze Estimation

Data augmentation (DA) is an effective technique and is widely used in various deep learning tasks (including gaze estimation). Appearance-based gaze estimation aims to directly learn a mapping from face images to gaze directions. Since subtle changes in eye regions are important for gaze estimation, direct data augmentation on faces is likely to damage key features in the eye region. We propose a facial landmarks based region-level data augmentation method. The method use facial landmarks to divide the face into eye regions and non-eye regions. Then we generate face images under different data augmentation methods by augmenting non-eye regions. We preserve the key features of eye regions. And the features of non-eye regions are augmented. We conduct experiments on the largest 2D dataset – GazeCapture. Comprehensive experiments show that the proposed method achieves promising results. Wide range of gaze estimation based application will be aspired from this work.