On Jul 23, 5:03 am, murat aykut <murat.ayk...@[EMAIL PROTECTED]
> wrote:
> Hi,
> I want to learn verification systems. For this purpose, I have
> searched a lot of papers (most of them about biometrics). But, I have
> some questions:
> - For the evaluation of FAR and FRR, must be all of the DB used? (In
> some papers leave one out like systems used, in anothers train-test
> like systems used)
You need enough training design data to accurately
estimate model parameters.
You need enough validation design data to estimate
errors precisely enough to make a choice between
competing candidate models.
You need enough nondesign test data to precisely
estimate the generalization errors (FAR and FRR
on unseen data).
If the data base is sufficiently large, randomly
choose what you need and be sure to include a
measure of the precision of your result.
If the data base is not sufficiently large,
consider resampling methods like bootstrapping
or cross-validation.
> - If I have a distance metric like Euclidean distance (I don't know
> max and min value), how can transform it to the matching score?
I am guessing that you have a two-class classifier
(C1,C2) with discriminant z = D2^2 - D1^2 and you
want to estimate false acceptance (assign C2 input
to C1) and false rejection (assign C1 input to C2)
rates based on the forced classification rule
z >= z0 assign x to C1
z < z0 assign x to C2
Use validation data to integrate the piecewise constant
z-histograms of each class to obtain the corresponding
piecewise linear CDFs. Plot 100*CDF1 (S-shaped) and
100*(1-CDF2) (reverse S-shaped) vs z. The curves yield
error rates for a given value of z. Where they intersect,
z0EQ, is the equal error rate.
Obtain a ROC curve from the CDF graph by plotting FRR(z)
vs FAR(z). Choose a suitable operating point on the curve
to determine z0.
If none of the points on the curve are suitable, you
need to use a two threshold conditional-classification
rule:
For z02 < z01,
z >= z01 assign x to C1
z < z02 assign x to C2
z01 and z02 can be chosen by specifying the desired
FRR and FAR.
Hope this helps.
Greg
|
