Can SIFT algorithm extract feature in real-time speed on PC?

Question

In page 25 of David Lowe's paper "Distinctive Image Features from Scale-Invariant Keypoints" in 2004, he claims, "Their computation is efficient, so that several thousand keypoints can be extracted from a typical image with near real-time performance on standard PC hardware." Here is the link: http://www.cs.ubc.ca/~lowe/papers/ijcv04.pdf

However, I tested SIFT algorithm using Andrea Vedaldi's sift++(aka VLFeat), which is a c++ implementation, on a 640x480 image, and it cost 0.839 second to extract about 3000 keypoints from one image. My PC is Intel i7 2600k, with 16GB RAM. Here is the link of the code: http://www.vlfeat.org/~vedaldi/code/siftpp.html

To be honest, I think it's quite weird if SIFT can achieve real-time speed, since it has to extract so much keypoints from one image.

Does anybody knows how fast is SIFT on a modern PC?

Answer 1

May be this article provide some insight about your question...Article actually includes parallelized versiob of SIFT but performance with one core version is also given.

Paralellization of SIFT