Local discriminant regions using Support Vector Machines for object recognition

Visual object recognition is a difficult task when we consider non controlled environments. In order to manage problems like scale, viewing point or occlusions, local representations of objects have been proposed in the literature. In this paper, we develop a novel approach to automatically choose which samples are the most discriminant ones among all the possible local windows of a set of objects. The use of Support Vector Machines for this task have allowed the management of high dimensional data in a robust and founded way. Our approach is tested on a real problem: the recognition of informative panels.