Eye tracking for foveation video coding and simple scene description

This paper investigates the use of eye tracking to define the important objects in a scene, i.e. those focused by viewers. These objects are rich in information, such as human face. They should be coded with higher quality than unfocused ones as in foveated video coding, for which eye tracking can serve as a preliminary step. We show that human gaze points often cover only a very small area of the screen, even less than 10%, providing great opportunities to save much more bit rate. Gaze points from a number of subjects watching a scene can be represented compactly in the proposed box diagram, another contribution from this work, that can give simple description of the scene just in a one diagram.