SIFT Implementation Based on Parallel Computation

Image matching is the basis of object recognition, tracking and three-dimensional reconstruction. The SIFT algorithm based on scale space has been widely used for image matching due to its strong robustness. However, the performance of the original flow is difficult to satisfy the real-time requirements. Due to the great advantage of parallel computing, this paper modified the algorithm flow of SIFT algorithm. The experimental results show that the modified algorithm flow makes the computation faster but can find more effective feature points. Also, the experiments further prove the invariance to change of image scales and rotation angles still exists apparently.