high performance computing on graphics processing units: hgpu.org

Scale-invariant feature transform (SIFT) was an algorithm in computer vision to detect and describe local features in images. Due to its excellent performance, SIFT was widely used in many applications, but the implementation of SIFT was complicated and time-consuming. To solve this problem, this paper presented a novel acceleration algorithm for SIFT implementation based on Compute Unified Device Architecture (CUDA). In the algorithm, all the steps of SIFT were specifically distributed and implemented by CPU or GPU, accroding to the step’s characteristics or demandings, to make full use of computational resources. Experiments showed that compared with the traditional implementation of SIFT, this paper’s acceleration algorithm can greatly increase computation speed and save implementation time. Furthermore, the acceleration ratio had linear relation with the number of SIFT keypoints.