KAZE Features

KAZE Features is a novel 2D feature detection and description method that operates completely in a nonlinear scale space. Previous methods such as SIFT or SURF find features in the Gaussian scale space (particular instance of linear diffusion). However, Gaussian blurring does not respect the natural boundaries of objects and smoothes in the same degree details and noise when evolving the original image through the scale space. By means of nonlinear diffusion we can detect and describe features in nonlinear scale spaces keeping important image details and removing noise as long as we evolve the image in the scale space. We use variable conductance diffusion which is one of the simplest cases of nonlinear diffusion. The nonlinear scale space is build efficiently by means of Additive Operator Splitting (AOS) schemes, which are stable for any step size and are parallelizable. Accelerated-KAZE Features uses a novel mathematical framework called Fast Explicit Diffusion (FED) embedded in a pyramidal framework to speed-up dramatically the nonlinear scale space computation. In addition, we compute a robust Modified-Local Difference Binary (M-LDB) descriptor that exploits gradient information from the nonlinear scale space. A-KAZE obtains comparable results to KAZE in some datasets, while being several orders of magnitude faster. Our results reveal a big improvement in repeatability and distinctiviness, for common 2D image matching applications