Local Image Descriptors for Matching and Classification Local Image Descriptors for Matching and Classification Title: Local Image Descriptors for Matching and Classification Acknowledgements

This thesis considers view-based object recognition in images, a problem that is still lacking an effective solution despite decades of research in computer vision. It is worth remembering that recognising an object from its appearance is considered as a keystone for image understanding, one of the most challenging problems in computer vision. The study and the analysis of the visual information coming from an image can be tackled with different approaches: to global image description we preferred the local approach since recent research has demonstrated that it leads to a more compact and robust representation of the image even when there are major changes in the object appearance. Thus in our work we concentrated on the use of local features and interest points to determine a representation of the image content by means of its most informative elements. We model 3D objects using a visual vocabulary whose words represent the most meaningful component of the object: the description obtained is complete and compact and is capable to describe the object when it is seen from different points of view. The robustness of this approach is remarkable also when the object is in a very cluttered scene and it is partially occluded. In respect to local approaches to object recognition, we focused on the following problems: • detection and description of local image features • estimation of the similarities between feature descriptors and matching points between images • formalisation of the concept of visual vocabulary and setting up of a classifier capable to compare several models of different objects. Within this framework, the contributions of this thesis are in the following areas: Matching techniques. We propose a matching strategy based on the use of a spectral method in association with local descriptors: since the representation we use is robust to scale and illumination changes, we obtain a compact and simple algorithm that can be used for severe scene variations. Object recognition. We present a method for 3D object modelling and recognition which is robust to scale and illumination changes, and to viewpoint variations. The object vocabulary is derived from a training image sequence of the object and the recognition phase is based on a SVM classifier. We introduced another adaptive strategy for object recognition in image sequences which is strongly based on the use of spatio-temporal coherence of local descriptors. In this case our work is motivated by the fact that an image sequence does not just carry multiple instances of the same scene, but also information on how the appearance of objects evolves when the observation point changes smoothly.