Hessian-Affine detector with SIFT descriptor


This is a modified version of Hessian-Affine detector proposed by Krystian Mikolajczyk. The initial scale selection is based on scale-space extrema of the Hessian operator. As it was shown by Lindeberg, Hessian operator has similar scale-selection properties as Laplace operator, thus with the Hessian we can simultaneously localize both position and scale of scale-space extrema (unlike in the Hessian-Laplace version proposed by Mikolajczyk). The affine adaptation procedure (also known as Baumberg iteration) is then applied on the scale covariant "Hessian-Hessian" points. The gravity vector assumption is used to fix the rotation uncertainty. Then, SIFT descriptors are computed on the affine normalized patches with measurement region of radius r = 3*sqrt{3}*s, where s is the scale of the point.

haff_cvpr09 image_name.ppm

The binary rewrites output file: image_name.ppm.hesaff.sift.

The output format is compatible with the binaries available from the page "Affine Covariant Features". The geometry of an affine region is specified by: u,v,a,b,c in a(x-u)(x-u)+2b(x-u)(y-v)+c(y-v)(y-v)=1. The top left corner of the image is at (u,v)=(0,0). The geometry of an affine region is followed by N descriptor values (N = 128).

File format:

u1 v1 a1 b1 c1 d1,1d1,2 d1,3 ... d1,N
um vm am bm cm dm,1dm,2 dm,3 ... dm,N

Related work: