Multi-CPU/Multi-GPU Based Framework for Multimedia Processing

Image and video processing algorithms present a necessary tool for various domains related to computer vision such as medical applications, pattern recognition and real time video processing methods. The performance of these algorithms have been severely hampered by their high intensive computation since the new video standards, especially those in high definitions require more resources and memory to achieve their computations. In this paper, we propose a new framework for multimedia (single image, multiple images, multiple videos, video in real time) processing that exploits the full computing power of heterogeneous machines. This framework enables to select firstly the computing units (CPU or/and GPU) for processing, and secondly the methods to be applied depending on the type of media to process and the algorithm complexity. The framework exploits efficient scheduling strategies, and allows to reduce significantly data transfer times thanks to an efficient management of GPU memories and to the overlapping of data copies by kernels executions. Otherwise, the framework includes several GPU-based image and video primitive functions, such as silhouette extraction, corners detection, contours extraction, sparse and dense optical flow estimation. These primitives are exploited in different applications such as vertebra segmentation in X-ray and MR images, videos indexation, event detection and localization in multi-user scenarios. Experimental results have been obtained by applying the framework on different computer vision methods showing a global speedup ranging from 5 to 100, by comparison with sequential CPU implementations.