An algorithm for associating the features of two images

Proceedings of the Royal Society B: Biological Sciences - Tập 244 Số 1309 - Trang 21-26 - 1991
G. Scott1, H. C. Longuet–Higgins
1Department of Engineering Science, University of Oxford, U.K.

Tóm tắt

Từ khóa


Tài liệu tham khảo

10.1098/rstb.1976.0090

Strang G. 1988 Linear algebra and its applications 3rd edn Appendix A. San Diego: Harcourt Brace Jovanovich.

10.1007/BF02424350

maximizes P:Gisnow the problem of finding the 1:1 Ullman S. 1979 The interpretation of visual motion ch. 2 and

mapping from is to js that minimizes the sum of the 3. Cambridge Massachusetts: M .I.T. Press.

squared distances r^between the members of each pair. Wertheimer M. 1912 Experimentelle Studien uber das

But if the positions of the circles and the crosses are

indeed connected by an affine transformation with a

positive definite tensor component then the mapping

between them does indeed minimize the sum of the

squares of the distances. So at last we can see why first

of all the P matrix delivered by our algorithm

approximates to a permutation matrix; and secondly

why the permutation represented by the matrix

faithfully recovers the 1:1 mapping originally induced

by the affine transformation.

D IS C U S S IO N In its classical formulation the correspondence

problem has to do with human vision and the

parameters that describe motion perception. But it is

also a problem for computer vision engineers and this

is how we have chosen to address it. In the cir

cumstances it is reassuring to find that our own

solution to the problem has a certain similarity to

Ullman's minimal mapping scheme a theory that

accounts for so many observations in the area of

motion perception. There are however important

differences. Like us Ullman aims to extremize the inner product

of two matrices: one having to do with `affinities I -or

`costs ' -and the other a matrix of correspondences. He

has much to say about such measures though he does

not explicitly consider the Gaussian proximity matrix

devoting much of his discussion to a cost measure that

is linear in the separation. This leads him to minimize

where possible the sum of the first powers of the

distances between corresponding points rather than

the sum of their squares which we find to be Sehen von Bewegung. Z. Psychol. 61 161-265.