Face Localization - Hierarchical Neural Networks for Image Interpretation - page 211

sides of the input images were wrapped around to fill the missing pixels. The moving
input was presented to the network that was trained with static images, without
any modifications. The network was able to move the output blobs along with the
moving input. Thus, it tracks the eyes. Figure 10.16 shows the relative distance d eye
over time. After few iterations, the distance reaches a level of about 0 . 075 . It varies
around this value until the end of the sequence. Interesting are the steep drops after
iterations 40 and 120, where the direction of movement is reversed. Here, the blobs
catch up with the movement. Hence, the output blobs follow the input motion with
a short delay.
10.5 Conclusions
In this chapter, an approach to face localization was presented that is based on the
Neural Abstraction Pyramid architecture. The network is trained to solve this task
even in the presence of complex backgrounds, difficult lighting, and noise through
iterative refinement.
The network's performance was evaluated on the BioID dataset. It compares fa-
vorably to a hybrid reference system that uses a Hausdorff shape matching approach
in combination with a multi-layer perceptron.
The proposed method is not limited to gray-scale images. The extension to color
is straight forward. Since the network works iteratively, and one iteration takes only
a few milliseconds, it would also be possible to use it for real-time face tracking by
presenting image sequences instead of static images. It was demonstrated that the
network is able to track a moving face.