2019/05/31 by Murad Qasaimeh, Kristof Denolf, Qasaimeh, Murad +8 · 6 citations
Engineering · Computer Science · #CCD and CMOS Imaging Sensors #Advanced Neural Network Applications #Advanced Image and Video Retrieval Techniques
paper · pdf · doi:10.48550/arxiv.1906.11879
Developing high performance embedded vision applications requires balancing\nrun-time performance with energy constraints. Given the mix of hardware\naccelerators that exist for embedded computer vision (e.g. multi-core CPUs,\nGPUs, and FPGAs), and their associated vendor optimized vision libraries, it\nbecomes a challenge for developers to navigate this fragmented solution space.\nTo aid with determining which embedded platform is most suitable for their\napplication, we conduct a comprehensive benchmark of the run-time performance\nand energy efficiency of a wide range of vision kernels. We discuss rationales\nfor why a given underlying hardware architecture innately performs well or\npoorly based on the characteristics of a range of vision kernel categories.\nSpecifically, our study is performed for three commonly used HW accelerators\nfor embedded vision applications: ARM57 CPU, Jetson TX2 GPU and ZCU102 FPGA,\nusing their vendor optimized vision libraries: OpenCV, VisionWorks and\nxfOpenCV. Our results show that the GPU achieves an energy/frame reduction\nratio of 1.1-3.2x compared to the others for simple kernels. While for more\ncomplicated kernels and complete vision pipelines, the FPGA outperforms the\nothers with energy/frame reduction ratios of 1.2-22.3x. It is also observed\nthat the FPGA performs increasingly better as a vision application's pipeline\ncomplexity grows.\n