Implementing interactive 3D segmentation on CUDA using graph-cuts and watershed transformation

Abstract

In this paper we present a novel scheme for a very fast implementation of volumetric segmentation using graph cuts. The main benefit of this work is our approach to non-grid region adjacency processing on CUDA which to our knowledge has not been done yet in any efficient way. The watershed transform radically reduces the number of vertices for graph processing. Everything starting from watershed transformation and ending with graph cut was parallelized and is performed directly on the GPU.