This award funds investigation of an approach to detecting and recognizing deformable articulated shapes, such as people and animals, from real images. The approach involves specifying a "language" for representing such objects, based on previous work on deformable template models. These models represent objects probabilistically, where the probabilities are required to specify likely geometric deformations of the object and of the imaging function. To handle effects such as lighting changes or partial occlusion of the object. These models, however, are complex, and so detecting them in images poses formidable combinatorial optimization challenges. This research explores a strategy which makes use of a small number of image "primitives" which can be thought of as key features. These may include both local image structures such as image corners and larger structures such as symmetric shapes. These should be simple enough that well-established optimization techniques can detect them optimally. The primitives are desaribed probabilistically, like the object models, enabling the use of information theory to quantify how much information a particular primitive provides about the preseance or absence, or configuration of the object in an image. It also enables optimal choices of primitives to be made, with the goal of finding primitives that are easy to compute and convey maximal information about the objects in the image.
