high performance computing on graphics processing units: hgpu.org

This paper spotlights the maintaining of scene boundary detection system in video and process of porting it to the OpenCL. The scene boundary detection algorithm proposed by authors is based on bottom-up focus attention principle. The system builds Gaussian pyramids from input image, calculates map of saliency from the image and then detects the most salient regions. The scene cut is detected when correlation of previous and next image’s histograms of salient areas is lower than threshold. To decrease the algorithm’s running time a parallel architecture was implemented. Several tricks and techniques which served for reaching better performance are described.