Image interpolation for DIBR viewsynthesis using graph fourier transform

Given texture and depth maps of one or more reference view-point(s), depth-image-based rendering (DIBR) can synthesize a novel viewpoint image by mapping texture pixels from reference to virtual view using geometric information provided by corresponding depth pixels. If the virtual view camera is located closer to the 3D scene than the reference view camera, objects close to the camera will increase in size in the virtual view, and DIBR´s simple pixel-to-pixel mapping will result in expansion holes that require proper filling. Leveraging on recent advances in graph signal processing (GSP), in this paper we propose to select appropriate graph Fourier transforms (GFT)-adaptive to unique signal structures of the local pixel patches-for expansion hole filling. Our algorithm consists of two steps. First, using structure tensor we compute an adaptive kernel centered at a target empty pixel to identify suitable neighboring pixels for construction of a sparse graph. Second, given the constructed graph with carefully tuned edge weights, to complete the target pixel we formulate an iterative quadratic programming problem (with a closed form solution in each iteration) using a smoothness prior in the GFT domain. Experimental results show that our algorithm can outperform in-painting procedure employed in VSRS 3.5 by up to 4.57dB.