3D videos have witnessed a growing interest in the last few years. Due to the recent development ofstereoscopic and auto-stereoscopic displays, 3D videos provide a realistic depth perception to the user and allows a virtual navigation around the scene. Nevertheless, several technical challenges are still remaining. Such challenges are either related to scene acquisition and representation on the one hand or to data transmission on the other hand. In the context of natural scene representation, research activities have been strengthened worldwide in order to handle these issues. The proposed methods for scene representation can be image-based, geometry based or methods combining both image and geometry. In this thesis, we take advantage of image based representations, thanks to the use of Multi-view Video plus Depth representation, in order to preserve the photorealism of the observed scene, and geometric based representations in order to enforce the compactness ofthe proposed scene representation. We assume the provided depth maps to be reliable.Besides, the considered cameras are calibrated so that the cameras parameters are known but thecorresponding images are not necessarily rectified. We consider, therefore, the general framework where cameras can be either convergent or parallel. The contributions of this thesis are the following. First, a new volumetric framework is proposed in order to mergethe input depth maps into a single and compact surface mesh. Second, a new algorithm for multi-texturing the surface mesh is proposed. Finally, we address the transmission issue and compare the performance of the proposed modeling scheme with the current standard MPEG-MVC, that is the state of the art of multi-view video compression.
