Determining Linear Shape Change: Toward Automatic Generation of Object Recognition Programs

Abstract

A 3D object localization task may be divided into two steps. First, one object appearance will be classified into one of the topologically equivalent classes of the 3D object appearances, referred to as aspects of the object (aspect classification). Then, the precise attitude and position of the object will be determined within one aspect (linear shape change determination (LSCD)). We have been working on designing a compiler which automatically generates an object localization program from a given object and sensor model; the compiler adheres to the above two-step strategy of aspect classification and linear change determination. For the first step, the compiler generates a decision tree structure program. Each branch node of the tree represents a neceSSary decision, such as comparing the area size of a visible face, to classify one object appearance into a smaller number of aspect groups. Along this tree, an object appearance is to be classified into one particular aspect at a leaf node of the tree. This paper will investigate the design of the compiler component to generate the second step. The compiler extends each leaf node of the tree and connects several nodes so that it performs the LSCD. The compiler chooses the largest 3D face as the primal face among several visible faces at the aspect corresponding to a leaf node. The compiler has priority rules which will select one particular method out of several possible ones; this rule defines a face coordinate system on a primal face. By using these rules, it analyzes the primal face, defines the face coordinate system on it, and registers the defining method to a node connected to the leaf node. The compiler also embodies the transformation from the face to the body coordinate system at the node. In order to increase the accuracy of the attitude and position, the compiler further puts two more nodes at each branch of the program. The first node establishes correspondences between model edges and image edges. The second node iteratively solves the transformation equation to determine the object's attitude and position using these correspondences. We have prepared a program library, which is a collection of prototypical objects to perform tasks mentioned before. In compile mode, the compiler retrieves necessary strategies defined at each node, finds proper prototypical objects from the program library, instantiates and inserts these objects to the node. In run mode, these instance objects run and determine the object's attitude and position through message-passing between objects.