high performance computing on graphics processing units: hgpu.org

This paper presents a new GPU-based tensor voting implementation which achieves significant performance improvement over the conventional CPU-based implementation. Although the tensor voting framework has been used for many vision problems, it is computationally very intensive when the number of input tokens is very large. However, the fact that each token independently collects votes allows us to take advantage of the parallel structure of GPUs. Also, the good computing power of modern GPUs contributes to the performance improvement as well. Our experiments show that the processing time of GPU-based implementation can be, for example, about 30 times faster than the CPU-based implementation at the voting scale factor sigma = 15 in 5D.