Detecting pairs of eyes

A strong characteristic of the eyes is that they often look dark in images. This feature is used to quickly find potential eyes in a image by looking at dark area in the image. Once a dark area is located, the parameters of the eye models around the area are computed. We can then estimate the probability of the area being a eye or not.

This search provide us with a lot of potential candidates. The best one is selected by looking at the relative positions of pairs of eyes. All selected eyes are tested by pairs and the probability of being a correct pair of eyes is computed from the distance and the angle given by each pair of eyes. The best match is then found using these probabilities. This gives use the most likely pair of eyes in a face images.

An active appearance model algorithm can then be applied to find the face more accurately. Indeed, given that we have the position of the eyes, we can use this information to find a good initialization for the AAM search algorithm. This initialization can be done by rotating, scale and translating the mean shape learnt from the training set so that it fits the pair of eyes found. Figure 4.2 shows how a shape can be aligned with eyes .

Figure 4.2: Alignment of a shape to fit eyes. Figure 4.2(a) shows the mean shape of a AAM model. Figure 4.2(b) shows how the mean shape can be rotated, translated and scaled to fit the positions of the eyes.

The search strategy is performed for every single frame of the sequence.