Abstract
Traditional multi-view stereo reconstruction via volumetric graph cuts formulates the 3D reconstruction problem as a
computationally tractable global optimization using graph cuts. It benefits from a volumetric scene representation and
discrete photo consistency is defined on the edge cost with a weighted graph. As the independence between each discrete
voxel, it is natural to do the parallel processing with multi-core CPUs or GPU, but after the photo consistency has been
estimated, it still need to design a parallel optimized methods to get the optimized labeling results for each voxel. In our
paper, we use the parallel volumetric graph cuts methods to solve the above problems. Our algorithm has two main steps,
clustering step and parallel graph cuts optimization step. We also introduce an approach for enhancing accuracy and
speeding up existing Multi-view 3D reconstruction methods, which based on volumetric graph cuts. The main idea is to
decompose the collected photos into some overlapping sets, while the voxels are also be clustered. The voxels
consistency estimating and surface labeling with graph cuts are processed in parallel, however, the labels of the
overlapped voxels may in general have multiple label solutions. It will be constrained to be equal to obtain a unique
solution in parallel graph cuts optimization step.
