Determination of Face Position and Pose with a Learned Representation Based on Labeled Graphs Determination of Face Position and Pose with a Learned Representation Based on Labeled Graphs

We present a neural system for the automatic determination of the position, size and pose of the head of a human gure in a camera image. The system is based on Elastic Graph Matching. The aspect of a human head in a speciic pose in an image is modeled by a labeled graph. The nodes of the graph refer to landmarks centered on speciic points of a head, e.g., the tip of the nose. Edges between nodes are labeled with distance vectors. The shape of a landmark is encoded in the form of a jet. A jet is a set of numbers extracted from the image by a family of wavelets that are all centered on one image point. The nodes of a pose graph are labeled by a bunch of diierent jets taken at the same landmark in diierent portraits. We therefore speak of \bunch graphs." A pose is determined by matching a number of bunch graphs that model diierent poses to the image and picking the one matching best. We here introduce a pose determination system which is based on well known face recognition system published previously. The pose determination system is characterized by a certain reliability and speed. We improve this performance and speed with the help of statistical estimation methods. In order to make these applicable, we reduce the originally very high dimensionality of our system with the help of a number of a priori principles. The rst of these is locality in the form of independence of diierent landmarks. This reduces model space to a set of independent subspaces. We reformulate the given global quality criteria as principles that can be applied to individual nodes and optimize the system according to these principles. We achieve signiicant improvement in speed and reliability. We discuss a possible extension of the learning algorithm aiming a completely autonomous object recognition system at the end of the paper.