Color FERET人脸数据库说明

The Color FERET Database

1. Contact

Stan Janet
National Institute of Standards and Technology
[e-mail]

1. Introduction

This database contains a total of 11338 facial images.
They were collected by photographing 994 subjects at various angles,
over the course of 15 sessions between 1993 and 1996. The Color FERET
website is:
http://www.nist.gov/humanid/colorferet

This database is largely a color version of the original Facial Recognition
Technology (FERET) Database:
http://www.nist.gov/humanid/feret/
which was released in 20001 and consisted of 14051 grayscale images.
In this documentation, the original FERET Database will be referred to as
the Gray FERET Database.

The database was created to develop, test and evaluate face recognition
algorithms. To advance the state-of-the-art in face recognition, the
Color FERET Database is made available to researchers in face recognition
on a case-by-case basis only. All requests for the Color FERET Database
must be submitted in writing, along with a signed copy of a release
agreement, available at:
http://www.nist.gov/humanid/colorferet/release_agreement.pdf

The images in the Color FERET Database are 512 by 768 pixels, and the
files are in PPM-format:
http://astronomy.swin.edu.au/~pbourke/dataformats/ppm/
Many image viewers can read PPM-format files. The PBMPLUS toolkit contains
software to convert PPM files to various other image formats, and it can be
downloaded free from:
http://www.acme.com/software/pbmplus/

To allow distribution of the database on a pair of DVD’s, the PPM files
have been compressed (losslessly) with the open source program “bzip2”:
http://sources.redhat.com/bzip2/
The program to decompress the files, “bunzip2,” is provided for several
platforms on the DVD’s, as is the “bunzip2” source code.

There are 13 different poses. (The orientation “right” means
facing the photographer’s right.)
fa regular frontal image
fb alternative frontal image, taken shortly after the
corresponding fa image
pl profile left
hl half left - head turned about 67.5 degrees left
ql quarter left - head turned about 22.5 degrees left
pr profile right
hr half right - head turned about 67.5 degrees right
qr quarter right - head turned about 22.5 degrees right
ra random image - head turned about 45 degree left
rb random image - head turned about 15 degree left
rc random image - head turned about 15 degree right
rd random image - head turned about 45 degree right
re random image - head turned about 75 degree right

Note that the rb, rc, and re angles are 5 degrees different from
the estimated values that were used in the Gray FERET Database
so that the angle increments would be more regular.

A regular frontal image was captured for every subject at every capture
session. In almost every case, a second frontal was also captured, and
half- and profile- left and right images were usually captured as well.

A breakdown of the images by pose is:
Pose Angle Images Subjects
fa 0 1364 994
fb 0 1358 993
hl +67.5 1267 917
hr -67.5 1320 953
pl +90 1312 960
pr -90 1363 994
ql +22.5 761 501
qr -22.5 761 501
ra +45 321 261
rb +15 321 261
rc -15 610 423
rd -45 290 236
re -75 290 236

The film was originally processed onto Kodak PhotoCD’s. The PhotoCD-
format images on those discs were converted into PPM-format using
the program “hpcdtoppm”:
http://www.mit.edu/afs/athena/contrib/graphics/src/netpbm-9.25/ppm/hpcdtoppm/

Since the amount of image data exceeds the capacity of a single DVD,
the data was split between the discs. Each subject who participated in
the collection had been given a unique five-digit ID. The imagery (and
ground truth files) for subject ID’s through “00739” are on the first
disc, and the others are on the second.

1. Summary of Differences

This database contains color versions of largely the same images that
appeared in the Gray FERET Database. There are four other significant
differences between this database and the Gray FERET Database:

1. The images' dimensions have been doubled to 512 x 768 pixels.2. The images have not undergone lossy compression or any other    processing. Compression or processing artifacts are very    visible in many Gray FERET images when comparing them to    their Color FERET counterparts. For example, mapping Color    FERET's 00863_940307_pl.ppm to grayscale and comparing    directly to Gray FERET's 00863pl010_940307.pgm shows that    the latter is much lower quality.    3. Numerous mistakes present in the Gray FERET Database have been    fixed.    4. After examining the images on the PhotoCD's, 96 images that    were never labelled and released have been added to this    database. The other 11242 originally appeared in the Gray    FERET Database. (Actually, a small number of those are new    images as well, where an image was erroneously in Gray FERET    under multiple names, and the error was fixed by manually    locating the image for the misnamed filename on the PhotoCD's.)

Two subsets of images from Gray FERET do not appear in Color FERET. The
2,200 FERET images known as the b-series (200 subjects each photographed
in 11 poses) are not present because the original PhotoCD’s could not be
located. Also, the 536 Gray FERET images that were created by manually
altering (scaling, darkening, or retouching) other images were not re-derived
for this database.

In addition, numerous mistakes discovered in the Gray FERET Database
have been fixed. Some of the types of mistakes were:
* an image appearing under multiple names
* a subject appearing under two different ID’s
* two different subjects appearing under the same ID
* incorrect capture date
* illegal filename (one Gray FERET file contained a hyphen)
The Gray FERET files affected by mistakes are detailed in the
file ‘doc/fates.out’, which is documented below.

The DVD’s also contain “smaller” (256 x 484) and “thumbnail” (128 x 192)
versions of each of the images, for applications that don’t require the
higher resolution. These images have also been compressed with “bzip2.”

There were four types of duplicates, photographs appearing under two or
more names, that were discovered on Gray FERET and fixed in Color FERET
by throwing out the extras. Some had the same checksum, and thus could
be found easily. The others had different checksums and fell
into three categories. In one, the Gray FERET TIFF-format files had
different checksums but the pixel values of each were identical, so
the checksums matched after they were converted to the PPM format.
In the second category, the images were visually similar, and the fact
that they were duplicates was ascertained because the differences
between each and the corresponding PhotoCD color image were both so
small. Finally, a few were fairly different visually because one
had been processed (e.g. darkened) – the duplicates were usually
found when the processed image was a close second-place to the
unprocessed image.

1. Meta-data

The database also contains ground truth files for the subjects and the
images, in both XML and plain text formats. This information allows
subsets of images, such as males, females, subjects with
glasses or beards, etc. to be used or analyzed independently.

The subject and recording XML ground truth files conform to document
type definitions (DTD’s) maintained at hbase.humanid.org:
http://hbase.humanid.org/hbase/dtd/recording.dtd
http://hbase.humanid.org/hbase/dtd/subject.dtd
To request access to the website, contact:
Todd Scruggs [e-mail]

To decompress the images, the program “bunzip2” is required. Source
code for “bunzip2” is provided on the DVD’s. It is usually standard
in Linux distributions, so there may be no need to compile the source
code.

Finally, the contents of the Gray FERET CD’s from 1999 are included
on the second Color FERET DVD, so that everyone with the Color FERET
Database has access to the previous database as well.

1. Image Naming Conventions

The full-size (512 x 768) images have names such as:
data/images/00012/00012_930831_fb_a.ppm.bz2
| | | |
| | | ______ optional flag
| | ________ FERET pose name
| _______________ the image capture date
______________________ the subject ID

The subject ID is a five-digit code that is unique for every
person photographed. The image capture date is in year-month-day
format. The 14 capture dates were:
930831
931230
940128
940307
940422
940519
940928
941031
941121
941201
941205
960530
960620
960627

For any of these three conditions, the filename will contain the
optional underscore and flag character:
a subject is wearing glasses
b subject changed his or her hairstyle for the
image and is not wearing glasses
c subject changed his or her hairstyle for the
image and is wearing glasses

The corresponding “smaller” and “thumbnail” images have names
such as:
data/smaller/00012/00012_930831_fb_a.ppm.bz2
data/thumbnails/00012/00012_930831_fb_a.ppm.bz2

1. Ground Truth Files

a. Subjects

There are plain text and XML-format files for each subject
containing sex, race, and approximate year of birth, all surmised
retrospectively. For example, information about subject ‘00012’
is in:
data/ground_truths/name_value/00012/00012.txt
data/ground_truths/xml/00012/00012.xml

There is also a single XML file containing that information
for all 994 subjects:
data/ground_truths/xml/subjects.xml

b. Recordings

There are plain text and XML-format file for each image
containing capture date, pose name, whether the subject was
wearing glasses or not, whether the subject has a beard or
mustache, and other information, e.g.:
data/ground_truths/name_value/00012/00012_930831_fb_a.txt
data/ground_truths/xml/00012/00012_930831_fb_a.xml

There is also a single XML file containing that information
for all 11338 images:
data/ground_truths/xml/recordings.xml

For approximately 90% of the Gray FERET frontal images, the
recording ground truth data contains the coordinates of the
subject’s mouth, nose and eyes. The data has been carried
forward to the color versions of those images. (It’s not
known why the remaining frontals were not truthed to have
those coordinates.)

1. Standard Testing Subsets

The directory ‘doc/partitions’ contains lists of frontal images that
constitute a gallery and three probe sets for the Color FERET Database:
doc/partitions/fa.txt
doc/partitions/fb.txt
doc/partitions/dup1.txt
doc/partitions/dup2.txt

They are the recommended sets for testing facial recognition performance
on the Color FERET Database. For each image in one of the probe sets,
the recognition task is to find the subject’s image in the gallery.

The gallery and probe sets are analogous to those used in the Gray
FERET tests of September 1996 and described in:
http://www.nist.gov/humanid/feret/feret_master.html
However, since the Color and Gray FERET Databases’ galleries and probe
sets are considerably different (as detailed below), there can be no
direct comparison between a system’s recognition performance on the
Color FERET sets and others’ on the Gray FERET sets. Some of the published
Gray FERET results are at:
http://www.itl.nist.gov/iad/humanid/feret/perf/eval.html
http://www.dodcounterdrug.com/facialrecognition/Feret/documents.htm

The Color FERET gallery consists the earliest ‘fa’ frontal image of
each subject. (If more than one image was captured on the earliest
date, the one with the subject not wearing glasses was selected.)
The gallery size is 994 images.

The ‘fb’ probe set is for analyzing the effect of a different facial
expression on recognition performance. It consists of the alternate
frontal (‘fb’) images taken just seconds after the gallery image.
The ‘fb’ probe set size is 992 images. (The subjects 00535 and 00989
did not have ‘fb’ images.)

There are two probe sets for analyzing the effect of aging on the
recognition performance. The first, ‘dup1’, consists of all other
frontals of the subjects. Most were captured at later sessions,
although some are images of the subjects at the earliest session
but with glasses and/or an alternate hairstyle. The ‘dup1’ probe
set size is 736 images.

The ‘dup2’ probe set is a subset of the ‘dup1’ probe set consisting
of all frontals captured at least 540 days after the subject’s gallery
image. The ‘dup2’ probe set size is 228 images.

There are several differences between the Color FERET and Grey FERET
galleries and probe sets:
1. The 194 subjects in the b-series that were in Gray FERET
sets are not present in Color FERET. Therefore the
Color FERET gallery and ‘fb’ probe sets each have
approximately 200 fewer images than their Gray FERET
counterparts.
2. The Gray FERET images that were derived from other images
are replaced in Color FERET with the untouched
originals.
3. Gray FERET contained some subjects erroneously under two ID’s
(see “doc/fates.out”). Those subjects each have a single
ID in Color FERET.
4. The Color FERET gallery contains images of all subjects in
the Color FERET Database. A few individuals in Gray
FERET were not represented in the Gray FERET gallery.

There are also 11 other files for each of the non-frontal poses:
doc/partitions/{[phq][lr],r[abcde]}.txt
They can be used as probe sets for analyzing facial recognition
performance when the subject is facing away from the camera to
various degrees.

The files ‘fa.txt’ and ‘fb.txt’ contain contain two columns: the first
is the subject ID and the second is the image basename. All of the
other files contain five columns. The first two columns are the same
as in ‘fa.txt’ and ‘fb.txt’. The third is the capture date of the image,
the fourth is the capture date of that subject’s gallery image, and the
fifth is the number of days between those dates.

1. Montages

The second DVD contains variety of montages. A subject
montage is an image that contains thumbnail images of all
images of the given subject. A pose montage contains all the
images in the database where the subject has a given pose.
Because these montages were so large, they’ve been divided
into several smaller montages, each with subjects of one
sex and race. Miscellaneous montages contain all images in
the database where the subject has or doesn’t have glasses,
a beard, or a mustache. These montages are also divided by
sex and race.

For each of subject montage, e.g.:
data/montages/id/00033.ppm.bz2
all images from the first capture date are on the top row.
The poses for each capture date are arranged on that row
left-to-right in the following order:

column  pose======  ====1   profile-left2   half-left3   quarter-left4   fa5   fb6   quarter-right7   half-right8   profile-right9   ra10  rb11  rc12  rd13  re

When there is no image for a certain pose on a given date, the
position in the montage is blank.

Images of the subject from the next capture date would be on
the next row. Images from the same capture date but which are
alternates (i.e. the subject was also photographed with glasses
or, in a few cases, with an alternate hairstyle) are also on
a separate row from the regular images.

For each pose, as well as for frontal images, there are montages
with names such as:
data/montages/pose/pr/M_Asian.ppm.bz2
data/montages/pose/frontals/F_White.ppm.bz2

Similarly, for glasses, there are montages with names such as:
data/montages/misc/no_glasses_F_Black-or-African-American.ppm.bz2
data/montages/misc/glasses_F_Black-or-African-American.ppm.bz2
And for beards and mustaches, there are montages of the males
of all races, both with and without them:
data/montages/misc/beard_M_Pacific-Islander.ppm.bz2
data/montages/misc/mustache_M_Pacific-Islander.ppm.bz2
data/montages/misc/no_beard_M_Pacific-Islander.ppm.bz2
data/montages/misc/no_mustache_M_Pacific-Islander.ppm.bz2

Empty montages were not be produced for combinations where
there were no representative images.

The file:
data/montages/partitions/partitions.ppm.bz2
is a montage of the Color FERET Database’s standard testing subsets. Each
subject starts a new row. The first and second columns of each row are
the subject’s ‘fa’ and ‘fb’ images. The rest of the row are ‘dup1’ images.
If the subject has any ‘dup1’ images that are also ‘dup2’ images, they’re
separated from normal ‘dup1’ images by a black divider image. For the few
subjects with more than 9 images, the rest are wrapped onto the
next line(s).

1. Additional Database Notes

Three of the PhotoCD’s contained grayscale images. Those 210 images
are listed in the file doc/grayscale.out.

Dozens of subjects were photographed in one session both with and
without glasses. Three female subjects – 00537, 00547 and 00556 –
were also photographed in the ‘940519’ session with a different
hairstyle and without glasses. The latter two were additionally
photographed again with the different hairstyle and with glasses.
The filenames for the alternate images contain the optional ‘b’
and ‘c’ flags as described previously.

1. Additional Files

doc/documentation.txt
This file.

doc/files.out
A file containing 11338 image basenames, starting with
00001_930831_fa_a.ppm, ending with 01208_940128_pr_a.ppm.

doc/aging.out
A file with a 5-column line for each of the 11338 images, e.g.:
00002 00002_931230_fa.ppm 931230 930831 121
which can be used to quickly determine the number of days
each image was captured after the subject’s first image
was captured. Column one is the subject ID, column two is
the image basename, column three is the image’s capture date,
column four is the capture date for the subject’s earliest
image, and column five is the number of days between those
dates.

doc/fates.out
Specifies what happened to each image in the old Gray
FERET Database. There are three main categories: 1) not
present (derived images, or “b-series” images that came
from PhotoCD’s that were not located); 2) present, with
no problems, so the new filename is straightforward; and
3) present, with some problem fixed, so the new filename
reflects the fix. The following are examples of many of the
varieties of ‘fate’ lines:

105 feret/cd1/data/pgm/00011fa010_930831.pgm : correct subject ID is 00012, wearing glasses; now named 00012_930831_fa_a.ppm

106 feret/cd1/data/pgm/00011fb010_930831.pgm : correct subject ID is 00012, wearing glasses; now named 00012_930831_fb_a.ppm

107 feret/cd1/data/pgm/00012fa010_930831.pgm : still present as 00012_930831_fa.ppm

108 feret/cd1/data/pgm/00012fb010_930831.pgm : still present as 00012_930831_fb.ppm

…

110 feret/cd1/data/pgm/00013fa001d_931230.pgm : derived images in original FERET Database were not re-derived

…

131 feret/cd1/data/pgm/00019ba010_960521.pgm : b-series PhotoCD’s of 05-21-96 were not located

…

779 feret/cd1/data/pgm/00086fa010_930422.pgm : thrown out; incorrectly-named duplicate of feret/cd1/data/pgm/00086fa010_940422.pgm; wrong date

781 feret/cd1/data/pgm/00086fa010_940422.pgm : correctly-named duplicate of thrown-out feret/cd1/data/pgm/00086fa010_930422.pgm

…

2107 feret/cd1/data/pgm/00247pr010_941121.pgm : incorrectly-named duplicate of feret/cd1/data/pgm/00247re010_941121.pgm; correct image located on PhotoCD’s

2112 feret/cd1/data/pgm/00247re010_941121.pgm : correctly-named duplicate of renamed feret/cd1/data/pgm/00247pr010_941121.pgm

…

2055 feret/cd1/data/pgm/00240fa010_940128.pgm : correct subject ID is 00256

…

4961 feret/cd1/data/pgm/00535fb010_940519.pgm : formerly designated as the ‘fb’ but there was no ‘fa’

…

9324 feret/cd2/data/pgm/00780qr010_950530.pgm : correct date is 960530

doc/gray_to_color_feret.map
A two-column file that maps the Color FERET images back to their
grayscale counterparts, basically a machine-readable version of
the above file, excluding the “not present” category:
…
feret/cd1/data/pgm/00012fa010_930831.pgm 00012_930831_fa.ppm
feret/cd1/data/pgm/00012fb010_930831.pgm 00012_930831_fb.ppm
feret/cd1/data/pgm/00011fa010_930831.pgm 00012_930831_fa_a.ppm
feret/cd1/data/pgm/00011fb010_930831.pgm 00012_930831_fb_a.ppm

doc/subject_counts.out
Contains the counts of images and frontal images for
each subject ID.
Column 1 is the subject ID
Column 2 is the image count
Column 3 is the frontal image count

doc/breakdown.out
Contains the image counts, frontal image counts, ‘fa’ image counts, and
subject counts for these subsets of the data:
1. females of each of the races represented
2. males of each of the races
3. each of the races
4. females
5. males
6. the entire Color FERET Database

doc/capture_dates.out
Contains a 3-column table. The image capture dates are in the first
column. The second and third columns contain the number of days between
the row’s capture date and, respectively, the first and the previous
capture dates.

doc/grayscale.out
Lists the 210 images that were originally
captured on black-and-white (actually grayscale) film. The red,
green and blue components for each pixel are identical.

binaries/bunzip2-102-x86-linux24
binaries/bunzip2-100-sparc-solaris7
binaries/bunzip2-100-x86-cygwinb20.exe
binaries/bunzip2-100-x86-win32.exe
binaries/bunzip2-102-x86-win32.exe
The program to decompress the image files, “bunzip2,” is provided
for several platforms
binaries/readme.txt
Installation information for “bunzip2” for those platforms

source/bzip2-1.0.2.tar
Source code for the “bzip2” package, which includes the “bunzip2”
program

gray_feret_cd1/
gray_feret_cd2/
[on DVD #2 only]
Contain the entire contents of the Gray FERET CD’s released in 2001.
The CD’s have been copied verbatim. Supplemental information and
errata can be found on the FERET website:
http://www.nist.gov/humanid/feret/

gray_feret_partitions.pgm.bz2
[on DVD #2 only]
A montage of the standard testing subsets for the Gray FERET Database.
The layout is the same as that of the montage of the Color FERET
Database’s standard testing subsets.

1. For More Information

The Color FERET Database’s main webpage is:
http://www.nist.gov/humanid/colorferet
and any supplements will be available there.

Any errata will be documented at:
http://www.nist.gov/humanid/colorferet/errata.html

Please direct comments, questions and bug reports to:
[e-mail]

1. Other URL’s of Interest

http://www.nist.gov/humanid/feret/feret_master.html
Gray FERET Database documentation

http://www.dodcounterdrug.com/facialrecognition/Feret/feret.htm
Department of Defense’s Counterdrug Technology Development Program FERET page;
the FERET program ran from 1993 through 1997 and was sponsored by the Defense
Advanced Research Products Agency (DARPA).

http://www.dodcounterdrug.com/facialrecognition/Feret/documents.htm
Gray FERET papers

http://www.frvt.org/FRVT2000/default.htm
http://www.frvt.org/FRVT2002/default.htm
The first and second Facial Recognition Vendor Tests, evaluations
of commercial face recognition systems

http://www.frvt.org/DLs/FR101.pdf
“Facial Recognition 101”