|Tipo di tesi||Tesi di dottorato di ricerca|
|Titolo||Dal Riconoscimento alla Segmentazione e ritorno|
|Titolo in inglese||From Recognition to Segmentation and return|
|Settore scientifico disciplinare||ING-INF/05 - SISTEMI DI ELABORAZIONE DELLE INFORMAZIONI|
|Corso di studi||Scuola di D.R. in INFORMATION AND COMMUNICATION TECHNOLOGIES (ICT)|
|Data inizio appello||2015-03-10|
|Disponibilità||Accessibile via web (tutti i file della tesi sono accessibili)|
In questa tesi sono affrontati due problemi centrali della visione artificiale: il riconoscimento e la segmentazione di oggetti. In entrambe le problematiche vengono utilizzate immagini del mondo reale, in cui i vincoli imposti (dal punto di vista alla forma e posa degli oggetti) andranno a definire la complessità del problema in esame.
In this thesis, two main problems in computer vision are analyzed: object recognition and object segmentation. Real world images are used in both settings where the imposed constraints (object dimensions, pose, number of objects, etc…) will define the complexity of the problem. Object recognition deals with recognizing the class of an unknown object among a set of possible categories, object segmentation has the goal of finding the contours of an unknown object and is usually used as a pre-processing step for further image understanding algorithms. A new image descriptor is proposed to enhance object recognition accuracy while introducing dataset independence in the computation of the image signature. The proposed descriptor is tested on large scale datasets and in various applications (from Cultural Heritage Imaging to Online Image Retrieval). In object segmentation, we first focus on class-specific supervised segmentation, where the category of the object to be segmented is known a priori, and it is used to build class-specific models. Object similarity is used in our proposal to identify similar training images and exploited by a One Class Support Vector Machine to find the appropriate segmentation model. When compared to a state-of-the-art proposal based on Structural Support Vector Machines, our method is able to obtain the same segmentation accuracy with a training procedure of one order of magnitude faster. A second proposal focuses on the graph-cut segmentation algorithm on superpixels (homogeneous groups of pixels) and introduces a learning scheme to find the borders of objects analyzing adjacent superpixels appearance. The last algorithm proposed in the thesis is about class-independent object segmentation, and investigates the connection between visual similarity and segmentation similarity, defined as the property of two objects of sharing the same segmentation challenges. A tree of segmentation prototypes is created at training time, formed by objects that share both visual and segmentation similarity. The visual appearance of an unknown object is used to search for the best prototype at testing time. The last chapter of the thesis is dedicated to applications of the proposed methods, where the ideas discussed above are tested to solve practical problems. All the methods proposed in this thesis are tested on publicly available datasets, to promote comparison with our works and for the sake of fairness.