Thèse soutenue

Compression multi-vues par représentation LDI (Layered Depth Images)
FR  |  
EN
Accès à la thèse
Auteur / Autrice : Vincent Jantet
Direction : Christine GuillemotLuce Morin
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2012
Etablissement(s) : Rennes 1
Ecole(s) doctorale(s) : École doctorale Mathématiques, télécommunications, informatique, signal, systèmes, électronique (Rennes)
Partenaire(s) de recherche : autre partenaire : Université européenne de Bretagne (2007-2016)

Résumé

FR  |  
EN

This thesis presents an advanced framework for multi-view plus depth video processing and compression based on the concept of layered depth image (LDI). Several contributions are proposed for both depth-image based rendering and LDI construction and compression. The first contribution is a novel virtual view synthesis technique called Joint Projection Filling (JPF). This technique takes as input any image plus depth content and provides a virtual view in general position and performs image warping while detecting and filling cracks and other small disocclusions. A pixel confidence measure is introduced to avoid ghosting artifacts in the rendered views. For intermediate view interpolation, JPF is used in collaboration with a floating texture realignment technique. For virtual view extrapolation, JPF is combined with a novelfull-Z depth aided inpainting technique. In order to efficiently encode the proposed LDI representation, a compression scheme basedon MVC/AVC standard is adapted to exploit both temporal redundancies and inter-layer redundanciesin the LDI sequence. An incremental construction scheme for the LDI is proposed,called I-LDI. This construction scheme reduces the completion rate of additional layers. Anobject-based layer organization of the LDI is then presented which ensures spatial consistency ofeach layer, and thus improves compression efficiency in comparison with a standard AVC/MVC scheme in rate-constrained context. Two rendering methods are finally proposed: the first one uses the JPF method, while the second one uses a 3D mesh for real-time rendering on aneight-views auto-stereoscopic display.