Learning models for object recognition

We consider learning models for object recognition from examples. Our method is motivated by systems that use the Hausdorff distance as a shape comparison measure. Typically an object is represented in terms of a model shape. A new shape is classified as being an instance of the object when the Hausdorff distance between the model and the new shape is small. We show that such object concepts can be seen as halfspaces (linear threshold functions) in a transformed input space. This makes it possible to use a number of standard algorithms to learn object models from training examples. When a good model exists, we are guaranteed to find one that provides (with high probability) a recognition rule that is accurate. Our approach provides a measure which generalizes the Hausdorff distance in a number of interesting ways. To demonstrate our method we trained a system to detect people in images using a single shape model. The learning techniques can be extended to represent objects using multiple model shapes. In this way, we might be able to automatically learn a small set of canonical shapes that characterize the appearance of an object.