Make Scale Invariant Feature Transform “Fly” with CUDA

Abstract

This paper introduces an implementation of scale invariant feature transform (SIFT) algorithm with CUDA. Primary  steps  including  building the  Gaussian  pyramid  and the difference of Gaussian pyramid, identification, localization [1], and orientation  generation of key-points  are  realized  on GPU with CUDA. A detailed description of important kernel function implementations is covered along with optimizations made to achieve high performance, and a comparison between the CUDA version SIFT algorithm and a baseline sequential CPU implementation is included.