A Memory Efficient Parallel Method for Voxel-Based Multiview Stereo

Voxel-based multi-view stereo is a method for reconstructing a 3D scene from a set of images taken at different viewpoints. This method reconstructs the 3D scene by defining a voxel space and determining the voxels corresponding with object surfaces. Due to the excessive processing time, some parallel methods have been proposed based on the voxel space division approach. However, a naive implementation of the existing parallel methods requires a large amount of memory and limits the use of a large scale voxel space, thus resulting in a low-quality reconstruction. To overcome this problem, this paper presents a memory-efficient parallel method for the voxel-based multi-view stereo. To reduce the memory usage in each processing element (PE), our method introduces a correspondence table which records the assigned PE and the location for each voxel. Experimental results show that although the processing time is slightly increased, the proposed method enables the reconstruction in large scale voxel space (up to 700×700×700) which is unattainable by the existing methods.