ABSTRACT
Comparing images to establish whether they depict the same object is an easy task for humans, who can perform it at a glance. On the contrary, although this problem has been investigated for many years, the results provided by the current computer vision systems are still poor (Pinto, Cox, and DiCarlo 2008). In Computer Vision, to compare image regions means matching features describing their visual appearance, like for instance color, shape, and texture. This is the core task of the object recognition challenge, i.e. posing the question of whether an image portion is an instance of a known object. In facing this challenge, the main difficulty is to model the wide variability under which an object can appear in a scene. For example, in different images, the same object could be rescaled, partially occluded, differently oriented and/or illuminated. Thus, the features used to describe the object should be invariant to the largest possible number of circumstances, so that a stable and efficient recognition becomes feasible.
