Evaluation of Face
Recognition Algorithms

DATA: Data Resources

The FERET Database

Much of the material on this web site is based upon studies that used the FERET data. The FERET data comes as a 2 CD set containing 14,051 images, of which 3,816 are frontal images. Most previous work on FERET data has focused attention on the frontal images. The FERET data is available for research purposes. The process for obtaining a copy of the FERET data is to request it from NIST. Details are posted on the

• The FERET Database website.

This site also contains important information about the images, including an overview of how the FERET database is structured.

Face Data Preprocessing: Normalization.

The FERET 1996/97 studies pre-processed the raw face data. This pre-processing is a critical to many of the algorithms distributed though this web site. The original code for preprocessing of the FERET data is distributed with the NIST data. However, we have used this code and found it to be somewhat problematic. CSU has developed a new preprocessing tool that performs all of the same basic preprocessing steps, but we are more confident in the reliability and generality of this new code. It is availalbe as part of the CSU Face Identification Evaluation System. You may read a bit more about normalization here and in the CSU Face Identification Evaluation System Users Guide.

Commonly Used Data Partitions

The FERET 1996/97 studies used several important subsets of FERET images. In particular, many of the results were reported for a common set of training images, gallery images, and four different probe sets. Here are links to the files that list the FERET images used in these sets.

One point of interest is the overlap between these sets, both in terms of people, and in terms of images.

This overlap information summarized on our Partitions Overlap Page.