1Carnegie Mellon University, Electrical and Computer Engineering, PA, USA
2Northeastern University, Electrical and Computer Engineering, MA, USA
3Georgia Institute of Technology, Computer Science, GA, USA
4Independent, China
5University of Maryland, Machine Learning, USA
*Corresponding Author: yuhongmo@cmu.edu
Online Published on 16 December, 2024.
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.
SIFT, CUDA, Parallelism
