IPL
Calibrated Color Image Database:
Natural Objects under CIE D65 and CIE A
About the Color Image
Database
The database consists of 130 calibrated color
images of natural
objects under calibrated illuminations:
- Calibrated images:
images are given in CIE XYZ tristimulus values
- Natural
objects imply complex textures, mutual illumination and shadows,
which
induce non-linear effects in the tristimulus values.
- Calibrated illuminations
include diffuse CIE D65 and
diffuse CIE A illuminant.
The
database is suitable for accurate studies on color image statistics,
chromatic adaptation in natural environments and color constancy.
|
Why a New Color Image
Database?
This color image database was collected in the context of a
chromatic adaptation study based on color statistics. In this context
the following facts are relevant:
- The proper way to describe the physical input to any
(artificial or human) visual system involves either (i) absolute
radiances, hyperspectral images, or (ii) absolute tristimulus images.
- Even if the spectral reflectance of objects is known, the
simple flat-Lambertian world assumption to estimate changes in
tristimulus values when changing the spectral radiance of the
illuminant is not valid in natural objects due to mutual illumination,
shadows and specular reflections. For the same reason, estimating the
reflectance by simply dividing the radiances by the spectrum estimated
from a white reference sample in an hyperspectral scene is not correct
either.
- Accurate
results on chromatic adaptation (as for instance, data on corresponding
pairs) are usually given under controlled illuminants (e.g. CIE D65 and
CIE A). In order to derive the psyhophysical behavior from color
statistics under these illuminants, a wide enough ensemble of natural
reflectances is required to have a continuous enough distribution of
samples in the tristimulus space.
Many of the available image color databases have the
following limitations related to the above issues:
- Uncalibrated data
(digital counts of conventional cameras instead of radiances or
tristimulus values).
- Spectroradiometric image databases do not generally include
the reflectance of the objects, but it has to be estimated using a
white reference sample. Even if the illuminant in some of the available
images is similar to the required illuminant (e.g. CIE D65 or CIE A),
the scene is not usually available under both illuminants. Even if you dont
impose such a restrictive requirement (same scenes under both
illuminants), but you just look for the same class of objects under
both illuminants, you usually find that the databases are not large
enough to ensure that the scenes under the desired illuminants include a wide enough set of similar reflectances.
- The above is also true for the available tristimulus image
databases.
In this situation, we decided to collect and make publicly
available this calibrated database
including a wide set of natural
objects
under a pair of calibrated illuminants.
|
Calibration
and
Experimental
Procedure
We used a Macbeth Executive ligth chamber (Macbeth Inc.)
equipped with standard CIE
D65 and CIE A illuminants and we took the CIE XYZ pictures using a
calibrated image colorimeter Lumicam1300 (Instruments and Systems Inc.)
in the configuration shown below.
Experimental Setting
|
 |
|
CIE
D65
Illumination
|
CIE A Illumination |
In every picture the exposure time for each filter (X,Y,Z)
was adjusted to avoid over or under exposition in order to ensure the
picture was taken in the optimal operating range of the camera.
Pictures with misregistered channels (due to motion) and scenes outside
the operating range of the camera (very dark and very bright regions in
the same scene) were discarded.
The accuracy of
illuminants and measurements was checked by taking pictures of 10
hues pages of the Munsell Book of Color. We checked the accuracy in
chromaticity by comparing the measured CIE xy chromaticities with those
computed from the known reflectances of the samples and the known
spectral radiance of the illuminants. In this case (flat matte samples)
we neglected geometrical factors and applied the flat Lambertian
assumption in the theoretical prediction. Experimental and theoretical
results are shown below. Luminance accuracy was
checked with a PR-670
SpectraScan spectroradiometer on the standard white background of the
pages of the Munsell Book of Color. The accuracy in luminance was roughly within the limits
provided by the manufacturer (~3%).
|
Organization of the
Database
The database includes 65 different scenes of natural objects
under two illuminants (130 images)
plus
10
scenes
displaying
different
hue
pages
of
the
Munsell
Book
of
Color
(20 images). The image
size is 1000x1280 pixels. The images are classified as follows:
|
Download Files |
 images_munsell_D
(.mat - 273Mb)
images_munsell_A
(.mat - 273Mb)
|
Calibration set (20 Munsell Images):
10 hue pages of the Munsell Book of Color (20 images), mainly matte
flat surfaces without mutual illumination. |
images_natural_I_D
(.mat - 1,36Gb)
images_natural_I_A
(.mat - 1,36Gb) |
Natural Objects I (100 images):
Scenes of colored textures with complex spatial geometry (non-flat
surfaces).
28 scenes of plants and flowers (56 images),
13 textured
natural materials (26 images),
9 samples of textured colored fabric (18
images). |
images_natural_II_D
(.mat - 416Mb)
images_natural_II_A
(.mat - 417Mb)
|
Natural
Objects
II
(30 images):
Scenes of office
material with simple spatial geometry (mainly flat surfaces including
bright objects). |
| readme_database
(2kb) |
How
to
load
the
database
in
Matlab:
See the sample file readme_database.m
on how to read the images. |
complete_database
(.rar - 4,07Gb)
|
Download the complete dataset (all the
above files):
|
Images (Matlab arrays of size 1000x1280x3) are stored in a Matlab structure in each of the
corresponding *.mat
files above. Images in the structure are sorted
according to the order in the pictures below. Chromatic diagrams with
all the colors in each set are also shown below. For your convenience we include a file to
load the images in Matlab.
|
Related
Papers
(database
citation)
- V. Laparra, S. Jiménez, G. Camps and J. Malo. Nonlinearities and Adaptation of Color
Vision from Sequential Principal Curves Analysis. Submitted for
publication 2012
|
Colorlab (The Matlab
Color Science
Toolbox)
If you found this database interesting, you may be probably
interested in accurate colorimetric computation in Matlab. Please check
this link for more information on COLORLAB:
The
Matlab
Color
Science
Toolbox!.
|
|
