Organizers: Edwin R. Hancock (Univ. of York, UK), Vittorio Murino (IIT, Italy), Marcello Pelillo (Univ. of Venice, Italy), Richard Wilson (Univ. of York, UK)
Duration: full day
Abstract: The presentation will revolve around two main themes, which basically correspond to the two fundamental questions that arise when abandoning the realm of vectorial, feature-based representations, namely: How can one obtain suitable similarity information from data representations that are more powerful than, or simply different from, the vectorial. How can similarity information be used in order to perform learning and classification tasks ? We shall assume no pre-existing knowledge of similarity-based techniques by the audience, thereby making the tutorial self- contained and understandable by a non-expert. The tutorial will commence with a clear overview of the basics of how dissimilarity data arise, and how it can be characterized as a prerequisite to analysis. We will focus in detail on the differences between Euclidean and non-Euclidean dissimilarities, and in particular the causes of non-Euclidean artifacts, how to test for them and when possible correct for them. With the basic definitions of dissimilarity to hand, we will move on to the topic of analysis in the dissimilarity domain, we will commence by showing how to derive dissimilarities for non- vectorial data, how to impose geometricity on such data via embedding and how to learn in the dissimilarity domain. Finally, we will illustrate how these ideas can be utilised in the computer vision domain with particular emphasis on the dissimilarity representation of shape.
Outline:
- Introduction to similarity-based pattern recognition
- Vector-space, distance and similarity
- Euclidean embedding techniques (standard methods, MDS etc)
- Non-Euclidean data (causes, tests, corrections)
- Non-Euclidean embedding techniques (spherical embeddings)
- Deriving similarities for non-vectorial data
- Hybrid generative/discriminative classification
- Generative kernels
- Information theoretic kernels
- Imposing geometricity on non-geometric similarities
- Structure preserving embeddings (random walks, zeta functions, path and cycle based methods)
- Complexity level characterizations of relational structures
- Algorithms on embedded dissimilarity data ( graph and hypergraph matching, clustering and feature selection)
- Learning with non-(geo)metric similarities
- Game-theoretic models of pattern recognition
- Polymatrix games and contextual pattern recognition
- Evolutionary games and data clustering
- Applications to problems in shape analysis (MRI image analysis, renal carcinoma diagnosis)
Tutorial material: http://simbad-fp7.eu/tutorialECCV2012.html