|M.Sc Student||Gattegno Gil|
|Subject||Object Recognition Using Geometric Hashing Extensions|
|Department||Department of Computer Science||Supervisors||Professor Ilan Shimshoni|
|Professor Michael Lindenbaum|
|Full Thesis text|
Geometric Hashing (GH) is a traditional object recognition technique, designed to recognize flat objects. The algorithm relies on affine coordinates, which, for flat objects, are pose invariant (under the weak perspective imaging model). The aim of this paper is to extend the GH method to recognizing non-flat objects in a scene.
Our approach is to preserve the original Geometric Hashing algorithm efficiency, while allowing recognition with various degrees of non-flatness. We propose two extensions of the Geometric Hashing algorithm that enables recognition of objects that are not necessarily flat (shallow objects). The two extensions enable a choice between a more reliable algorithm (2.5D Geometric Hashing) and a faster one (Extended Geometric Hashing).
Our extensions present two major enhancements to the traditional GH algorithm: (a) taking into account the object's non-flatness in calculating the range for the variation of the affine coordinates (b) using basis filtering to remove bases which are not likely to be good. While the Extended Geometric Hashing algorithm uses a general bound on the location error, resulted by the affine coordinates variation, the 2.5 Geometric Hashing uses a more detailed prediction, which specifies a consistency constraint.
We evaluated the performance of these algorithms using simulation and real data experiments. The 2.5 Geometric Hashing reaches a success rate of above 90% even when object’s non-flatness increases. Its reliability is kept at high 94% values, even when the clutter was three times larger. For shallow objects the Extended Geometric Hashing algorithm gives the same reliability but is faster, and is therefore preferred.