View-based object recognition using saliency maps

We introduce a new view-based object representation , called the saliency map graph (SMG), which captures the salient regions of an object view at multiple scales using a wavelet transform. This compact representation is highly invariant to translation, rotation, and scaling, and ooers the locality of representation required for occluded object recognition. We then introduce two graph similarity algorithms, one computing the topological similarity between two SMG's and the other computing the geometrical similarity between two SMG's. To reduce search complexity, the k model SMG's for a given object become the leaves in a binary tree, allowing the closest SMG (or view) to be found in O(log k) similarity computations. We test and compare these similarity and search algorithms in the context of a view-based object recognition problem involving a large database of model views.