The first part of his work focuses on the image representation problem and proposes compact saliency-driven image descriptors able to capture perceptually coherent structures independently of their position in the scene. Complementing the saliency-driven representations, Marco also proposed effective classification algorithms (ML3 and NBNL) able to provide complex decision boundaries at a fraction of the computational resources required by standard classifiers (e.g. Gaussian kernel SVM and NBNN).
Potential applications of his work include vision-based spatial reasoning for mobile robots and automatic organization of digital collections of images.
Specific topics of his thesis are: automatic visual scene recognition, bag-of-visual-words representations, saliency maps, feature pooling, locally-linear SVM, latent SVM, multi-component classification algorithms and naive Bayes nearest neighbor algorithms.
General topics of his thesis are: image classification, computer vision, pattern recognition and machine learning.
To download Marco's thesis, click on the following link: Saliency-based Representations and Multi-component Classifiers for Visual Scene Recognition