Estimating 3D Facial Pose in Video with Just Three Points

Abstract

Estimating facial pose along a video sequence is an important problem in many applications such as perceptual user interfaces, people monitoring, and aid for the disabled. Currently, most existing methods to compute facial pose are designed to obtain high accuracy under controlled conditions. However, efficiency and robustness are preferable in most practical situations. In this paper, we deal with the problem of 3D pose estimation with the aim to design a fast and approximate solution that requires a minimum amount of information: our method is able to work with just the location of three points -the eyes and the mouth- in each frame. Using that information, and introducing several as sumptions - that could hold in many cases-, we show how to estimate the six degrees of freedom of facial position and orientation. Moreover, one single point could be enough after initializing the system. The obtained accuracy greatly depends on the precision of the underlying 2D face tracker. The system is generic and does not require to be trained for each new user.