An algorithmic skeleton for massively parallelized mean shift computation with applications to GPU architectures

In this paper we discuss parallelization approaches for generic mean shift clustering. We provide an algorithmic skeleton which allows an easy creation of platform specific implementations, be it small scale systems as multicore CPUs, large GPUs or even distributed cluster systems. Additionally we provide an exhaustive runtime complexity analysis and various remarks for further research. In order to illustrate the practicability of our theoretic framework we discuss a GPU implementation which exhibits significant speedups for small and large scale datasets.