A floating polygon soup representation for 3D video

This thesis presents a new representation called floating polygon soup for applications like 3DTV and FTV (Free Viewpoint Television). The polygon soup is designed for compactness, compression efficiency, and view synthesis quality. The polygons are stored in 2D, with depth values at each corner. They are not necessarily connected to each other and can be deformed (or floated) w.r.t viewpoints and time. Starting from multi-view video plus depth (MVD the construction holds in two steps: quadtree decomposition and multi-view redundancy reduction. It results in a compact set of polygons replacing the depth maps while preserving depth discontinuities and geometric details. Next, compression efficiency and view-synthesis quality are evaluated. Classical meth- ods such as inpainting and post-processing are implemented and adapted to the poly- gon soup. A new compression method is proposed. It exploits the quadtree structure and uses spatial prediction. Results are compared with an existing MVD compression scheme based on MPEG?s H.264/MVC. A slightly higher PSNR value is obtained at medium and high bitrates and ghosting artifacts are greatly reduced. Finally, the polygon soup is floated according to the desired viewpoint. This view- dependent geometry is guided by motion estimation between synthesized and original views. This method reduces remaining artifacts and improves the final image quality.