>>

Soutenance de HdR : Hedi Tabia

Titre du mémoire

Contributions to 3D data analysis, retrieval and classification.

Date et lieu de soutenance

Mardi 21 novembre 2017, 14h.

ENSEA, salle du conseil.

Abstract

The work presented in my habilitation reviews the main results of my research activities carried out over the last five years. During this period, I have been particularly active in developing tools to describe, recognize, retrieve, and classify three dimensional (3D) Data. These are fundamental problems and building blocks to many applications in computer vision, computer graphics, medical imaging, and archeaology. The presented contributions goes around three main issues; namely 3D shape analysis, cross domain retrieval, and 3D action recognition. For 3D shape analysis, we proposed three main contributions. First, we proposed covariance matrices as new descriptors for 3D shape matching. Second, we proposed the use of a set of Riemannian metrics for computing distances between covariance matrices and analyzed their clustering behavior. Using a Riemannian metric on the manifold of covariance matrices, we introduced the concepts of Bag of Covariance matrices (BoC) and spatially-sensitive BoC, as extensions of the standard bag of words originally introduced in Euclidean spaces, for 3D shape retrieval and classification. We also extended the standard kernel methods, such as kernel Support Vector Machines (kSVM), to the space of covariance matrices for 3D shape analysis. For the cross domain retrieval, we addressed the issue of searching 3D shapes using different modalities. First, we proposed different view based 3D shape retrieval methods. Second, we proposed 3D shape retrieval method that uses 3D sketch as a query. Finally, we proposed a generic framework for multimodal shape querying that involves 3D shapes (3D meshes), 2D images (photos) and hand-drawn sketches. For 3D action recognition, we proposed different types of action features computed from the skeletons in the action sequences. We also proposed to learn feature combination in order to improve the recognition accuracy. We further proposed a multiple classifier combination system which offers the possibility to derive a measure of contradiction between the classifier decisions to be fused. This allowed us boosting the accuracy on the accepted data, which is desired in situations where a miss-classification is very expensive or must not happen. In this dissertation, I further present ongoing works and longer time perspectives on both shape and action analysis.

Composition du jury

  • M. Marc Pierrot-Deseilligny, Ecole Nationale des Sciences Géographiques
  • M. Robert B Fisher, University of Edinburgh
  • M. Frédéric Precioso, University Nice Sophia Antipolis
  • M. Tobias Schreck, Graz University of Technology
  • Mme. Michela Spagnuolo, National Research Council (CNR-IMATI)
  • Mme. Nicole Vincent, Université Paris Descartes
  • M. Dan Vodislav, University of Cergy Pontoise

Go back