EPFL doctorate Award 2010 - Kokiopoulou Effrosyni

"For her outstanding PhD thesis work about the geometry-based representation, analysis and classification of visual information sets."

Geometry-aware analysis of high-dimensional visual information sets.

Over the last fifteen years we have been experiencing a revolution in the amount of data that we collect and publish. This very fact creates a grand challenge: How can we create useful knowledge out of this data deluge? This question that refers to the problem of data analysis, lies at the heart of this thesis. We focus on the problem of pattern classification in the analysis of high-dimensional visual information sets that come from modern multimedia applications. High dimensions stem from the representation of data in high-dimensional vector spaces, such as the number of pixels of an image, the numerous frames of a video sequence or the several points in a 3D point cloud.

The thesis proposes new methods for the analysis of visual pattern manifolds based on sparse geometric expansions and graph models. The thesis' contribution is three-fold. First, it leverages the use of sparse representations towards invariance with respect to geometric pattern transformations as well as for classification-aware dimensionality reduction. The proposed method is guaranteed to find the globally optimal solution of the associated non-convex optimization problem, when the pattern transformation consists of a synthesis of translation, rotation and scaling. Second, it proposes graph-based methods for pattern classification with multiple observations in both centralized and distributed environments and third, it designs fast consensus methods for distributed data analysis and classification in multimedia sensor networks.