The same robustness, three times faster
In May 2006 Herbert Bay, Tinne Tuytelaars and Luc Van Gool presented SURF. Gaussian kernels are replaced by rectangular box filters computed in constant time through an integral image, the detector is built on the determinant of the Hessian, and the descriptor is cut to 64 numbers. On the same 800 by 640 image, detection plus description took 354 ms against 1036 ms for SIFT.
Why it matters
Invariant features became usable for things that have to answer at once. The cost of SIFT is exactly what is missing when working on video or on a device, and SURF showed the same robustness could be had three times cheaper and on a vector half as long.
The timing table on the first image of the Graffiti scene, 800 by 640, on a 3 GHz Pentium IV: U-SURF 255 ms, SURF 354 ms, SURF-128 391 ms, SIFT 1036 ms. A separate table for the detectors alone: Fast-Hessian 120 ms for 1418 points against 400 ms for difference-of-Gaussians, 650 ms for Hessian-Laplace and 1800 ms for Harris-Laplace. The applied experiment is recognising works of art in a museum: a reference database of 216 images of 22 objects and a test set of 116 images taken under extreme lighting changes, through the glass of display cases, from other viewpoints and at 320 by 240. Recognition rates: SURF-128 85.7 percent, U-SURF 83.8 percent, SURF 82.6 percent, GLOH 78.3 percent, SIFT 78.1 percent, PCA-SIFT 72.3 percent. Bay and Van Gool are given as being at ETH Zurich, Tuytelaars and Van Gool at the Katholieke Universiteit Leuven.