High performance computing model for action localization in video

Last decade, the field of High Performance Computing (HPC) has been rapidly developing. This advances in computational power and hardware resources motivated real-time video application. One of the most challenging application is “action detection” inside video. Beside its tremendous complication, it requires an intensive computation power for video analysis. This work extends our previous work in this problem by addressing both multi-core CPUs and Graphical Processing Unit (GPU) for implementing the video action localization system (previous work). In this paper, multiple (HPC) architectures are implemented to speedup real-time video streaming analytics. The experiments are conducted on the previously implemented recognition system which exploits Deep Learning models to localize an action inside a video. The results illustrates that CUDA implementation accomplished 8x speedup while CPU accomplished 5.2x speedup. The experiments were done on Sports-IM dataset with 487 action.