Ameba: A High-Performance and Energy-Efficient Online Video Retrieval System

This paper describes a high-performance and energy-efficient online video retrieval system called Ameba. Ameba contains a number of custom reconfigurable nodes which are grouped by a novel architecture. The system aims to address the performance and energy efficiency issues for large scale dynamic concurrent query requests. The open SURF approach and a Hamming distance based matching algorithm were implemented and improved on FPGA to increase the performance and energy efficiency. A predictive algorithm is proposed to forecast the trends of online query requests. This scalability which is obtained from the dynamic reconfiguration, makes the system have ability to improve its energy efficiency with the guarantee of performance. The simulation experiments are conducted with a considerable library which consists of 4650 videos with a combined length of more than 3000 hours. The comparative results demonstrate that Ameba has performance and energy efficiency advantages when facing large scale online video retrieval requests.