Guaranteed geometric hashing

Abstract

Geometric hashing is an invariant feature-driven approach to model-based object recognition. Previous interest has focused on its ability to accommodate sensor error. This paper presents an enhancement of the geometric hashing technique which guarantees, under only a few constraints, that models will not be missed due to sensor noise. The authors' geometric hashing algorithm enters model affine invariants into hash table regions defined by an exact error model, brings together known optimizations (table symmetry and the use of more than 3 model-scene point correspondences) and uses novel data organization. Experimental results (on both synthetic and real data) suggest that the authors' modifications to a geometric hashing recognition scheme effectively overcome sensor noise.