Performance optimizations for PatchMatch-based pixel-level multiview inpainting

As 3D content is becoming ubiquitous in today´s media landscape, there is a rising interest for 3D displays that do not demand wearing special headgear in order to experience the 3D effect. Autostereoscopic displays realize this by providing multiple different views of the same scene. It is however unfeasible to record, store or transmit the amount of data that such displays require. Therefore there is a strong need for real-time solutions that can generate multiple extra viewpoints from a limited set of originally recorded views. The main difficulty in current solutions is that the synthesized views contain disocclusion holes where the pixel values are unknown. In order to seamlessly fill-in these holes, inpainting techniques are being used. In this work we consider a depth-based pixel-level inpainting system for multiview video. The employed technique operates in a multi-scale fashion, fills in the disocclusion holes on a pixel-per-pixel basis and computes approximate Nearest Neighbor Fields (NNF) to identify pixel correspondences. To this end, we employ a multi-scale variation on the well-known PatchMatch algorithm followed by a refinement step to escape from local minima in the matching-cost function. In this paper we analyze the performance of different cost functions and search methods within our existing inpainting framework.