Multiview image compression based on LDV scheme

Benjamin Battin, Cédric Niquin, Philippe Vautrot et Laurent Lucas

In recent years, we have seen several different approaches emerge to deal with multiview compression. First, we can find the H264/MVC extension which generates quite heavy bitstreams when used on n-views autostereoscopic medias and does not allow inter-view reconstruction. Another solution relies on the MVD (MultiView+Depth) scheme which keeps p views (n > p > 1) and their associated depth-maps. This method is not suitable for multiview compression since it does not exploit the redundancy between the p views, moreover occlusion areas cannot be accurately filled. In this paper, we present our method based on the LDV (Layered Depth Video) approach which keeps one reference view with its associated depth-map and the n-1 residual ones required to fill occluded areas. We first perform a global per-pixel matching step (providing a good consistency between each view) in order to generate one unified-color RGB texture (where a unique color is devoted to all pixels corresponding to the same 3D-point, thus avoiding illumination artifacts) and a signed integer disparity texture. Next, we extract the non-redundant information into two textures (a unified-color one and a disparity one) containing the reference and the n-1 residual views. The RGB texture is compressed with a conventional DCT or DWT-based algorithm and the disparity texture with a lossless dictionary algorithm. Then, we will discuss about the signal deformations generated by our approach.