FRAMSTIM: framework for large scale multimedia content feature extraction based on MPI one-sided communication

Abstract

Every day a large number of images are made available throw social networks and different IoT embedded sensors. R&D devoted to the development of applications based on visual pattern recognition has attracted a large population of researchers in both side academic and industry. Extraction of relevant features is challenging and known to be one of the key issues in many applications where the visual pattern recognition is applied (object recognition and tracking, image identification, multimedia document categorization, indexing and retrieval, deep learning based visual feature coding, video surveillance, robotic, activity recognition). Furthermore the extraction features from a big volume of image and video data is time and resources consuming. In the context of the ITEA2 project H4H/PerfCloud ( Performance in the Cloud) we have developed parallel OpenMP threads video engine search. To scale the extraction of visual features from a large volume of streaming visual content, we have developed a framework based on OpenMP and MPI one-sided communication where the computation and communication are overlapped thanks to the RDMA approach.