Multimedia Mining on Manycore Architectures: The Case for GPUs

Media mining, the extraction of meaningful knowledge from multimedia content, poses significant computational challenges in today's platforms, particularly in real-time scenarios. In this paper, we show how Graphic Processing Units (GPUs) can be leveraged for compute-intensive media mining applications. Furthermore, we propose a parallel implementation of color visual descriptors (color correlograms and color histograms) commonly used in multimedia content analysis on a CUDA (Compute Unified Device Architecture) enabled GPU (the Nvidia GeForce GTX280 GPU). Through the use of shared memory as software managed cache and efficient data partitioning, we reach computation throughputs of over 1.2 Giga Pixels/sec for HSV color histograms and over 100 Mega Pixels/sec for HSV color correlograms. We show that we can achieve better than real time performance and major speedups compared to high-end multicore CPUs and comparable performance on known implementations on the Cell B.E. We also study different trade-offs on the size and complexity of the features and their effect on performance.