3. Facial Feature Detection

We have developed a system for facial features based on our informatio-based maximum discrimination detectors. Since the facial features are best described using edges, the discrete observation vector or image sub-window of each facial feature consists of the combination of three low-level image features: two intensity levels, vertical edges (3 levels) and horizontal edges (3 levels). Each element of this discrete observation vector has 18=2x3x3 possible values.

Using the face candidates found with our face/eye detector, a scale- and rotation-normalized feature image is obtained. Nine facial feature detectors are applied on this image. Examples of these low-level feature images and of the results of our nine feature detectors are shown in the following figure.