Computing OpenSURF on OpenCL and General Purpose GPU. (31st October 2013)

• Record Type: Journal Article
• Authors: Yan, Wanglong; Shi, Xiaohua; Yan, Xin; Wang, Lina
• Is Part Of: International journal of advanced robotic systems. Volume 10:Number 10(2013)
• Publication Date: 2013-10-31
• Subjects: OpenCL -- OpenSURF -- General Purpose GPU; Robotics -- Periodicals; Robotics; Periodicals; 629.892
• DOI: 10.5772/57057
• Languages: English
• ISSNs: 1729-8806

Abstract

Speeded-Up Robust Feature (SURF) algorithm is widely used for image feature detecting and matching in computer vision area. Open Computing Language (OpenCL) is a framework for writing programs that execute across heterogeneous platforms consisting of CPUs, GPUs, and other processors. This paper introduces how to implement an open-sourced SURF program, namely OpenSURF, on general purpose GPU by OpenCL, and discusses the optimizations in terms of the thread architectures and memory models in detail. Our final OpenCL implementation of OpenSURF is on average 37% and 64% faster than the OpenCV SURF v2.4.5 CUDA implementation on NVidia's GTX660 and GTX460SE GPUs, repectively. Our OpenCL program achieved real-time performance (>25 Frames Per Second) for almost all the input images with different sizes from 320*240 to 1024*768 on NVidia's GTX660 GPU, NVidia's GTX460SE GPU and AMD's Radeon HD 6850 GPU. Our OpenCL approach on NVidia's GTX660 GPU is more than 22.8 times faster than its original CPU version on Intel's Dual-Core E5400 2.7G on average.