Real-time visual hull computation based on GPU

A real time visual hull parallel computation method is proposed based on GPU. Firstly, an effective Octree data structure is built on GPU, which establishes the correspondence between voxels and the thread ID of CUDA. Secondly, for large-scale parallel data in the projection test of visual hull calculation, the convex hull of the projected vertices is accelerated to use the GPU programmable features. Finally, the redundant edge computation in marching cubes algorithm is reduced by building the edge lookup table. Then, parallel streaming data reduction based on prefix-sum excludes empty voxels from the lookup table, which can improve the computation effectively. Experimental results show that, for large-scale data, accurate visual hull can be efficiently reconstructed by the proposed algorithm on a common PC and GPU platform.