Unsupervised learning of a scene-specific coarse gaze estimator

Abstract

We present a method to estimate the coarse gaze directions of people from surveillance data. Unlike previous work we aim to do this without recourse to a large hand-labelled corpus of training data. In contrast we propose a method for learning a classifier without any hand labelled data using only the output from an automatic tracking system. A Conditional Random Field is used to model the interactions between the head motion, walking direction, and appearance to recover the gaze directions and simultaneously train randomised decision tree classifiers. Experiments demonstrate performance exceeding that of conventionally trained classifiers on two large surveillance datasets.