3D object recognition using invariant feature indexing of interpretation tables

Abstract

This paper proposes a technique for 3D object recognition from range data. Our technique (which was motivated by recent work in geometric hashing) employs precompiled interpretation tables as a source of prototype hypotheses. These tables are indexed by invariant features computed from triples of surface patches. We describe a method for automated construction of interpretation tables from CAD models. Invariant features extracted from the input image are used to index the tables, retrieve prototype hypotheses, and bind them to scene entities. These candidate hypotheses are then ranked via computation of a matching score. The use of precompiled tables of hypotheses offers the potential for greater efficiency in the recognition step than techniques such as constrained search. Experiments demonstrate significant reductions in the number of incorrect hypotheses produced over a constrained search mechanism.