Sj. Maybank, PROBABILISTIC ANALYSIS OF THE APPLICATION OF THE CROSS RATIO TO MODEL-BASED VISION - MISCLASSIFICATION, International journal of computer vision, 14(3), 1995, pp. 199-210
Citations number
14
Categorie Soggetti
Computer Sciences, Special Topics","Computer Science Artificial Intelligence
The cross ratio of four colinear points is of fundamental importance i
n model based vision, because it is the simplest numerical property of
an object that is invariant under projection to an image. It provides
a basis for algorithms to recognise objects from images without first
estimating the position and orientation of the camera. A quantitative
analysis of the effectiveness of the cross ratio in model based visio
n is made. A given image I of four colinear points is classified by ma
king comparisons between the measured cross ratio tau of the four imag
e points and the cross ratios stored in the model database. The image
I is accepted as a projection of an object O-sigma with cross ratio a
if \tau - sigma\ less than or equal to ntu, where n is the standard de
viation of the image noise, t is a threshold and u = \del tau\. The pe
rformance of the cross ratio is described quantitatively by the probab
ility of rejection R, the probability of false alarm F and the probabi
lity of misclassification p(sigma)(sigma), defined for two model cross
ratios sigma, sigma. The trade off between these different probabilit
ies is determined by t. It is assumed that in the absence of an object
the image points have identical Gaussian distributions, and that in t
he presence of an object the image points have the appropriate conditi
onal densities. The measurements of the image points are subject to sm
all random Gaussian perturbations. Under these assumptions the trade o
ffs between R, F and p(sigma)(sigma) are given to a good approximation
by R = 2(1 - Phi(t)), F = r(F) epsilon t, root p(sigma)(sigma) = e(si
gma)epsilon t\sigma - sigma\(-1), where epsilon is the relative noise
level, Phi is the cumulative distribution function for the normal dist
ribution, r(F) is constant, and e(sigma) is a function of a only. The
trade off between R and F is obtained in Maybank (1994). In this paper
the trade off between R and p(sigma)(S) is obtained. It is conjecture
d that the general form of the above trade offs between R, F and p(sig
ma)(sigma) is the same for a range of invariants useful in model based
vision. The conjecture prompts the following definition: an invariant
which has trade offs between R, F, p(sigma)(sigma) of the above form
is said to be non-degenerate for model based vision. The consequences
of the trade off between R and p(sigma)(sigma) are examined. In partic
ular, it is shown that for a fixed overall probability of misclassific
ation there is a maximum possible model cross ratio sigma(m), and ther
e is a maximum possible number N of models. Approximate expressions fo
r sigma(m) and N are obtained. They indicate that in practice a model
database containing only cross ratio values can have a size of order a
t most ten, for a physically plausible level of image noise, and for a
probability of misclassification of the order 0.1.