This thesis comes within the scope of the emerging 3D systems and their issues of multi-view-plus-depth coding, virtual view synthesis and human perception. The 3D stereoscopic perception is tackled through the proposal of a dynamic visual attention model, an efficient depth map coding method and a new rendering algorithm for viewpoint generation by extrapolation. The first part of this thesis focuses on the implications of the binocular disparity on the deployment of visual attention. After a statistical analysis of the role of potential center and depth biases in monoscopic and stereoscopic conditions, a new saliency model combining low and high level visual features is proposed. This dynamic saliency model integrates the hypothetical depth mechanism of figure/ground processing using depth foreground/background segregation. Tested performance proves the validity of the approach and confirms the relevance of a fusion of features whose weighting depends on time. The second part proposed two contributions in the scope of high quality 3D contents. A new depth map compression method based on lossless edge transmission provides a simple and reliable scene geometry for accurate viewpoint synthesis. This method is assessed thanks to objective quality metrics and subjective experiments. Finally a method of directional inpainting is presented. It is dedicated to extrapolation of new viewpoints for both 3DTV and FTV. The background structure is first propagated inside disoccluded areas. The robust tensor-based isophotes and the directional filling enable an efficient synthesis of virtual views. This provides promising visual results even for distant generated viewpoints.
