Parallel Partitioning: Path Reducing and Union–Find Based Watershed for the GPU

Abstract

The watershed transform is a common step in different image processing tasks. With the fast development of generalpurpose computing on GPUs, a number of parallel watershed algorithms have been introduced for that setup. We propose two novel parallel watershed algorithms: one intended for relatively larger images and another for smaller images. Our algorithms are based on the combination of the parallel path reduction technique, introduced in our previous paper on the topic, and the parallel version of the Union-Find algorithm. We show through multiple experiments, both in 2D and 3D, that our algorithms achieve unmatched execution times. On a typical gaming GPU, our algorithm processes an 800 megavoxel image in around 2.5 seconds.

Parallel Computation of Piecewise Linear Morse-Smale Segmentations

“…Algorithms based on topographic distance compute shortest paths to find corresponding catchment basins [43,46]. Gabrielyan et al [47] and Yeghiazaryan and Voiculescu [48] provide GPU implementations using an approach similar to path compression, speeding up the shortest path computation. Simulated immersion approaches seed catchment basins at minima, extending them by processing vertices in increasing function value order [44,49].…”