Accelerating mean shift image segmentation with IFGT on massively parallel GPU

Mean shift algorithm is a popular technique in many machine vision applications including image segmentation. Main drawback of the original algorithm is its quadratic computational complexity, the problem approached with many acceleration methods developed by researchers so far. One of the most effective is usage of the Improved Fast Gauss Transformation (IFGT) to accelerate Gaussian summations of the mean shift, resulting with linear computational complexity. Despite such advances, mean shift segmentation on larger images can still be too expensive for time critical applications. However, recent rapid increase in the performance of general purpose graphic processing unit (GPGPU) hardware has opened opportunity for significant acceleration of the algorithms by parallel execution. This paper introduces first parallel implementation of IFGT-MS segmentor based on many core GPGPU platform. The emphasis is placed on adaptation of the core algorithm to efficiently exploit benefits of underlying GPU hardware architecture. Numerical experiments have demonstrated considerably faster segmentation execution compared with alternative CPU and GPU based mean shift variants.