FERET Evaluation

Last Updated: June 07 2001  


This page documents the evaluations of fourteen of the algorithms tested under the FERET program that ran from 1993 to 1997. They were tested using defined partitions of the FERET database.

Two performance metrics are quoted:

Identification cumulative match scores
Verification receiver operating characteristics

The C++ code used to generate these results may be downloaded here.


A probe-by-probe analysis is given here:

Rank of match for each probe

The effect of gallery size is given here:

Identification Performance vs. Gallery Size

The algorithms reported on here are:

arl_corArmy Research LaboratoryBaseline Correlation
arl_efArmy Research LaboratoryBaseline Eigenface
ef_hist_dev_angNIST Seven eigenface-based systems a common image processing front end and eigenface representation but differing only in the distance metric: "ang" is angle; "anm" is something; "l1" is Manhattan distance; "l2" is euclidean; "ml1" and "ml2" are Mahalanobis weighted versions of "l1" and "l2"; and "md" is unknown.
ef_hist_dev_anmNIST
ef_hist_dev_l1NIST
ef_hist_dev_l2NIST
ef_hist_dev_mdNIST
ef_hist_dev_ml1NIST
ef_hist_dev_ml2NIST
excaliburExcalibur Inc
mit_mar_95MIT Media Lab
mit_sep_96MIT Media Lab
umd_mar_97University of Maryland
usc_mar_97University of Southern California

Four specific investigations are reported here:

dupI Duplicate One Images taken upto three years after their matching image
dupII Duplicate Two Images taken between 18 and 36 months after their matching image
fafb Expression Images of subjects with alternate facial expressions
fafc Illumination Images taken under different lighting conditions