Common Measure in Face Verification

In face verification task, we have several metrics to measure model performances.

Verification Rate (FR) under False Acceptance Rate

First, we need to understand the false acceptance rate. There is a table showing reality and prediction.

false acceptance rate is defined as “the percentage of identification instances, in which unauthorised cases are incorrectly accepted”.

$$ FAR = \frac{FA}{FT} = \frac{False_Negative}{False_Negative+True_Negative} $$

Verification rate is the accuracy of a verification model.

$$ VR = \frac{TP + TN}{all} $$

AUC means area under the curve. ROC is the Receiver operating characteristic.

x axis is False Positive Rate: the percentage of error in the positive reality $$ FPR = \frac{False_Positive}{False_Positive+True_Negative} $$

y axis is True Positive Rate: the accuracy in the positive reality $$ TPR = \frac{True_Positive}{True_Positive+False_Negative} $$

The closer the AUC is to 1, the better the model is.

To pick a good threshold of the model, it should pick the low FPR and high TPR point.