Gaze Estimation Based on Deep Learning Method

Many mature methods of gaze estimation are available in various scenarios. Relying on additional hardware or platforms with professional equipment to tackle intensive computation tasks is a prominent problem of traditional methods, which usually involves high costs and is relatively tedious. Besides, the implementation of traditional gaze estimation method is typically complex. Traditional gaze estimation approaches require systematic prior knowledge or expertise for practical operations, and the gaze is estimated through the representation of pupil and iris, so high-quality images shot in special environments are required. This paper proposes a data-driven method for gaze estimation. It can be applied to various mobile platforms with deep learning methods instead of additional hardware devices or systematic prior knowledge. When collecting gaze data set, the paper designs a set of automatic and fast data collection mechanism on the mobile platform. Beyond that, the paper proposes an annotation method on collected gaze dataset that improves the predicted accuracy. The results demonstrate that the deep learning method performs well and can satisfy the task need of different applications.