Re-sampling and interpolation of DIBR-synthesized images using graph-signal smoothness prior

In depth-image-based rendering (DIBR), a new virtual viewpoint image is synthesized by mapping color pixels from one or more reference views to the new image grid using corresponding disparity values. However, due to necessary roundings to the 2D grid, "rounding holes" appear in synthesized objects, which are typically filled using local interpolation. In this paper, leveraging on a recent 3D image compression scheme called graph-based representation (GBR) that losslessly codes disparity information, we propose instead to re-sample and interpolate missing on-grid pixels of an object at decoder using mapped off-grid color pixels as reference. Specifically, we first describe the underlying data kernel for the desired signal using a weighted graph, where the edge weights reflect non-integer distances between reference and missing pixels. A graph-signal smoothness prior is then assumed to complete missing pixel values via iterative unconstrained quadratic optimization. Experimental results show that the synthesized objects have better quality than conventional local interpolation methods.