Generic Recognition of Articulated Objects through Reasoning about Potential Function

Abstract A recognition system driven by the functional capabilities of objects should categorize objects according to their possible uses. For example, if an object provides an "opposing finger grasp," "opposing cutting blades," and a "closable grasp that causes cutting," then it would be classified under the category "scissors." Previous work in this area has not dealt with objects whose function depends on parts joined by articulated connections. This paper describes a function-based recognition system that can reason about articulated objects. The system is composed of two major stages. In the first stage, an articulated shape model is recovered from a sequence of observed shape instances. In the second stage, the articulated shape model is analyzed to determine its functional potential. The initial implementation of the system recognizes objects in the category "scissors." Experimental results are presented for 24 different object shapes, with varying degrees of membership in this category.