Parallelization of voxel based multiview stereo for arbitrarily configured viewpoints
Masaru Fukushi, Naoki Sekiguchi, Toru Abe · 2012
Voxel based multiview stereo is a method for reconstructing a 3D scene from a set of images taken at different viewpoints by determining the voxels corresponding to object boundaries in a voxel space. To accelerate the voxel based multiview stereo, we propose a novel parallel method based on the voxel space division approach, which divides the voxel space into several domains and assigns them to multiple Processing Elements (PEs). The existing methods based on this approach cannot work well for viewpoints arranged in complex configurations, since their process decomposition increases communication between PEs. Unlike those methods, we take new process decomposition, where the calculation for every voxel is decomposed into its component for each viewpoint. Our method processes the calculations for each of two viewpoints in parallel applying the voxel space division, and then integrates their results. This method can reduce the amount of inter-PE communication for arbitrarily configured viewpoints.