Visual Learning from Multiple Views

Abstract

An algorithm is presented in which a computer is visually shown a sequence of views of a solid planar object as the object is rotated in space. The computer automatically forms a three-dimensional description of the object. The description consists of a deterministic description of the object's surfaces and how they are interconnected to form the object, along with a measure of each surface's shape which is invariant to three-dimensional rotation. From this self-learned model of the object, the object can later be recognized from any viewing angle. The basis of the algorithm is the ability of the program to determine in a specific visual view: "What do I see now that I have seen before?" This is accomplished by generating two sets of mappings of one object description to another object description.