Using scale-space theory and the local shape of auto-correlation matrix

a new general approach is proposed to extract affine invariant regions.First

the feature points and their characteristic scales are detected by the local maxima of normal Gaussian derivatives over scale-space;Then

the auto-correlation matrices

which are used to describe local image structure

are computed on the characteristic scales of feature points.The extracted ellipse regions are affine invariant.Four affine invariant region algorithms

namely Harris3D

Laplace3D

Hessian3D and Localjet43D

are presented using the general approach.The experimental results show the four algorithms are invariant to illumination

rotation and scale changes.The affine invariance is verified by the simulation test we designed for affine invariance.Comparing the four algorithms

we find that the performance of other three algorithms is similar except for Harris3D.