5. Summary

In this chapter, the head pose estimation, one of the most challenging tasks in the area of computer vision, is introduced, and the main types of methods are demonstrated and compared. Particularly, the manifold learning-based methods are attracted more attention. In reality, data distribution is usually nonlinear in high-dimensional represented space, e.g., the head pose images. Some potential structures are lying on nonlinear but smooth manifolds which are embedded in the original space. The manifold learning algorithms are able to discover and visualize such embedding. Almost all the algorithms are formalized based on the assumption of the local linearity of the nonlinear data. Those algorithms highly benefit the application of head pose estimation, because the face orientations (yaw and pitch) are found to be distributed along some specific manifolds. Promising performance is achieved by the classical manifold learning methods, which, however, are highly improved by the supervised manifold learning. It proves that the supervised information represented as angles of head poses is helpful in head pose estimation. However, there are still hurdles to take. Most of the methods are tested in different settings, e.g., different database is used in different method. A common framework could help to offer fair justifications. Other feature instead of the simple color space can be considered to better represent the face images. Some useful tools are available online to help better understand the work [16, 32, 33].
