Gaze estimation using local features and non-linear regression

In this paper, we present an appearance-based gaze estimation method for a head-mounted eye tracker. The idea is to extract discriminative image descriptors with respect to gaze before applying a regression scheme. We employ multilevel Histograms of Oriented Gradients (HOG) features as our appearance descriptor. To learn the mapping between eye appearance and gaze coordinates, two learning-based approaches are evaluated : Support Vector Regression (SVR) and Relevance Vector Regression (RVR). Experimental results demonstrate that, despite the high dimensionality, our method works well and RVR provides a more efficient and generalized solution than SVR by retaining a low number of basis functions.