File sd4feats.tar.Z is a unix-compressed tarfile.  Uncompressing, then
untarring it produces directory sd4feats, containing text files
train.dat and test.dat.  File train.dat contains training feature
vectors made from the 2000 first-rolling fingerprints of NIST Special
Database 4, and file test.dat contains testing feature vectors made
from the 2000 corresponding second-rolling fingerprints of that
database.  Along with the feature vectors, these files contain target
vectors indicating the correct classes of the fingerprints.  These
files are being provided for use in testing pattern classification
programs.  Each feature vector consists of 112 floating point numbers,
made by a feature extractor that ends with the Karhunen-Loeve
transform.  Details about the feature extraction process used may be
found in:

  C. L. Wilson, G. T. Candela, P. J. Grother, C. I. Watson, and R. A.
  Wilkinson.  Massively Parallel Neural Network Fingerprint
  Classification System.  Technical Report NIST IR 4880, National
  Institute of Standards and Technology, July 1992.

  G. T. Candela and R. Chellappa.  Comparative Performance of
  Classification Methods for Fingerprints.  Technical Report NIST IR
  5163, National Institute of Standards and Technology, April 1993.

The above reports are publically available on the anonymous FTP server
sequoyah.nist.gov in the directory /pub/nist_internal_reports, as
compressed PostScript files.  The file README in that directory contains
a comprehensive listing of all reports there.

Documentation for Special Database 4, and for other databases
published by the Image Recognition Group, is also available on
the FTP server in the directory /pub/databases/manuals. The
Special Database 4 documentation is in the file fingerdoc.ps.Z.
Special Database 4 is too large to make available on the FTP
server; it is available on CD-ROM.  For information about ob-
taining Special Database 4 or other Image Recognition Group
databases, phone NIST Standard Reference Data at 301/975-2208.

The format of files train.dat and test.dat is as follows.  The first
two lines of each file are header information; following lines consist
of the Karhunen-Loeve feature vector, then the target vector
(indicating correct class), for each fingerprint.
  Line 1:
    N: Number of examples in the file (N = 2000)
    n: Dimensionality of the features contained in the file (n = 112)
    L: Number of classes (L = 5)
  Line 2:
    L strings.  Text labels for the L classes.  (Abbreviations A, L,
      R, T, W for classes Arch, Left Loop, Right Loop, Tented Arch,
      Whorl.)
  Lines 3 and onwards:
    For each of the N examples:
      n floating point values of the leading K-L coefficients
      L floating point target values.  The position of the 1.000000
	indicates which class the pattern belongs to. In all other
	positions there is a 0.000000.  For example, if a fingerprint
	is a Left Loop, then this is class 2, and the target
	vector for this fingerprint is therefore:
		0.000000 1.000000 0.000000 0.000000 0.000000

If you have any questions about train.dat and test.dat, contact:

G. T. Candela
email: jerry@magi.nist.gov
phone: (301) 975-3388
mail : Image Recognition Group
       Building 225, Room A-216
       National Institute of Standards and Technology
       Gaithersburg, MD 20899