High-performance computing for vision

Vision is a challenging application for high-performance computing (HPC). Many vision tasks have stringent latency and throughput requirements. Further, the vision process has a heterogeneous computational profile. Low-level vision consists of structured computations, with regular data dependencies. The subsequent, higher level operations consist of symbolic computations with irregular data dependencies. Over the years, many approaches to high-speed vision have been pursued. VLSI hardware solutions such as ASIC's and digital signal processors (DSP's) have provided good processing speeds on structured low-level vision tasks. Special purpose systems for vision have also been designed. Currently, there is growing interest in using general purpose parallel systems for vision problems. These systems offer advantages of higher performance, sofavare programmability, generality, and architectural flexibility over the earlier approaches. The choice of low-cost commercial-off-theshelf (COTS) components as building blocks for these systems leads to easy upgradability and increased system life. The main focus of the paper is on effectively using the COTSbased general purpose parallel computing platforms to realize high-speed implementations of vision tasks. Due to the successful use of the COTS-based systems in a variety of high performance applications, it is attractive to consider their use for vision applications as well. However, the irregular data dependencies in vision tasks lead to large communication overheads in the HPC systems. At the University of Southern California, our research efforts have been directed toward designing scalable parallel algorithms for vision tasks on the HPC systems. In our approach, we use the message passing programming model to develop portable code. Our algorithms are specified using C and MPI. In this paper, we summarize our efforts, and illustrate our approach using several example vision tasks. To facilitate the analysis and development of scalable algorithms, a realistic computational model of the parallel system must be used. Several such models have been proposed in the literature. We use the General-purpose Distributed Memory (GDM) model which is a simple but realistic model of state-of-theart parallel machines. Using the GDM model, generic algorithmic techniques such as data remapping, overlapping of communication with computation, message packing, asynchronous execution, and communication scheduling are developed. Using these techniques, we have developed scalable algorithms for many vision tasks. For instance, a scalable algorithm for linear approximation has been developed using the asynchronous execution technique. Using this algorithm, linear feature extraction can be performed in 0.065 s on a 64 node SP-2 for a 512 × 512 image. A serial implementation takes 3.45 s for the same task. Similarly, the communication scheduling and decomposition techniques lead to a scalable algorithm for the line grouping task. We believe that such an algorithmic approach can result in the development of scalable and portable solutions for vision tasks. © 1996 IEEE Publisher Item Identifier S 0018-9219(96)04992-4.published_or_final_versio