30 May 2006
Meritxell Vinyals

Object Recognition is one of the more important applications in Computer Vision and although there are a lot of literature about it the problem has not yet been solved . Finding good representations of the object invariant to a wide range of variations (like illumination changes, occlusion or different points of view ) is the main problem to achieve this goal. During the last few years, there has been a growing interest in the approach based on describing an object using local descriptors computed from local interest regions. This approach is more robust to occlusion, severe lighting changes or complex background than global descriptors. It will allow us to recognize and match an object with its model and estimate the transformation ocurred between them. In this talk, an object recognition's implementation based on local descriptors will be introduced with its three main phases: * Region detectors. Regions described need to be regions which are covariant to a wide range of variations so a region detector is needed. Although there are a lot of different detectors in the literature we'll give a brief overview and we 'll go on to focus on one of them: Extremas in DOG's (Difference-of-Gaussian) * Region descriptors. Once regions have been detected, they need to be described. Region descriptors will describe regions providing distinctive and invariance to variations. Usually, the choice of this descriptor will be independent from the region detector previously chosen. We'll focus on SIFT (Scale Invariant Feature Transform) descriptor which has showed quite good performance in the last reviews. The recognition will be reached by finding correspondences between these descriptors. * Pose Estimation methods. Finally, different pose estimation methods will be described in order to provide the accurate transformation between the model and the image.