A parallelization method for multiview stereo

This paper proposes a parallelization method for 3D shape reconstruction with voxel-based multiview stereo. In our approach, a target voxel space is divided into several domains, each of which is assigned to a different PE (processing element) to be processed in parallel. To reduce the amount of inter-PE communication, our proposed method divides the voxel space by planes passing through a virtual viewpoint computed from a configuration of actual viewpoints. Moreover, to reduce the idle time of PEs, our method performs dynamic load balancing flexibly, based on the estimated effect of it. Experimental results indicate that our method can reduce the processing time by 35% at 32 PEs as compared to a naive parallelization method.