VOIDGA: A View-Approximation Oriented Image Database Generation Approach

Figure 1.Comparison between the generated views (left and right) and the ground truth (middle) for the asteroid impact dataset yA31 at cycle time 29945. The orange and blue surfaces are contours of the temperature (0:2 eV) and water density field (0:002 g/cm3), respectively. The views have been approximated using only 24 depth images with a resolution of 5122 pixels, whereas the original triangulation has around 3 million triangles. Our approach first computes the position of each depth image pixel in world coordinates, and then renders the resulting geometry either via a point cloud (left) or a surface triangulation (right). The used depth images have been chosen with VOIDGA to bound the maximum approximation error for the current view.In this work, we propose a novel view-approximation oriented image database generation approach (VOIDGA) that enables the adequate generation of arbitrary views. Our approach utilizes Depth Image Based Rendering (DIBR) techniques to derive novel views based on a set of depth images. In contrast to approaches that store a huge amount of images to cover a wide range of possible view directions, VOIDGA identifies and stores only those images that significantly contribute to the overall view-approximation quality while bounding the resulting approximation error. This further reduces the size of image databases and the number of images that need to be processed by DIBR algorithms. We demonstrate VOIDGA on several challenging real-world examples, and compare our approximations against ground truth renderings using two image-based metrics.