The Recognition of Material
Properties for Vision and Video Communication
MIT2000-02
Progress Report: January 1,
2002ŃJune 30, 2002
Edward Adelson
Project
Overview
How
can we tell that an object is shiny or translucent or metallic by looking at
it? Humans do this effortlessly, but machines cannot. Materials are important
to humans and will be important to robots and other machine vision systems. For
example, if a domestic cleaning robot finds something white on the kitchen
floor, it needs to know whether it is a pile of sugar (use a vacuum cleaner), a
smear of cream cheese (use a sponge), or a crumpled paper towel (grasp with
fingers).
An
objectÕs appearance depends on its shape, on the optical properties of its
surface (the reflectance), the surrounding distribution of light, and the
viewing position. All of these causes are combined in a single image. In the
case of a chrome-plated object, the image consists of a distorted picture of
the world; thus the object looks different every time it is placed in a new
setting. Somehow humans are good at determining the "chromeness" that
is common to all the chrome images.
Progress
Through June 2002
Reflectance
Estimation
In
the last six months, we developed a theoretical foundation describing the
relationship between illumination statistics and the statistics of an image of
a surface with a particular reflectance.
This allows us to select statistics for reflectance classification,
using the framework described in previous reports. We also continued our psychophysical experiments to
determine which statistics humans use in order to estimate the reflectance of
objects.
Recovering
Intrinsic Images from Single Images
Estimating
material properties is difficult because an image is the combination of the
objectÕs material properties and other visual characteristics, such as the
shape of the objects and their illumination. We are developing a system that can isolate two of these
characteristics, the reflectance and shading of the objects in a scene. We use the term shading to refer to the
interaction of the objects shape and the illumination of the scene.
Previously,
we had developed a system to decompose a color image into an image representing
the shading of the scene and an image representing the reflectance of the
scene. In the last six months, we
have extended that system to gray-scale images. This is necessary because color information alone is not
sufficient to distinguish shading changes from reflectance changes. Figure 1 shows an example of such an
image. Using color information
alone, our system would incorrectly classify the mouth and eyes that have been
painted on the pillow as shading.
Figure 2 shows the results of just using color. However, by combining gray-scale and
color information, the shading and reflectance images are recovered correctly,
as shown in Figure 3.
The
system works by distinguishing image derivatives caused by a reflectance change
from those caused by shading. The
shading and reflectance images can then be recovered from the classified
derivatives. Derivatives in
gray-scale images are classified as shading or reflectance derivatives by using
filters that can discriminate patterns in the image that tend to be created by
shading from patterns which tend to be created from reflectance changes. These filters capture the different
statistics of shading and reflectance and are be found using a machine learning
technique based on the AdaBoost algorithm. Figure 4 shows another example.
We
have also developed an algorithm to improve the results in areas of the image
where it is not clear whether the image changes are caused by a reflectance
change or shading. This algorithm
propagates information from areas where the correct answer is clear into areas
where the correct answer is unclear.
Figure 1 - Both gray-scale and color
information is necessary to separate shading changes from reflectance changes
in this image.
|
|
Shading |
Reflectance |
Figure 2 - The shading and reflectance images generated by using color alone.
|
|
Shading Image |
Reflectance Image |
Figure 3 - The results obtained using
both color and gray-scale information.
|
|
|
Original Image |
Shading Image |
Reflectance Image |
Figure 4 Š An example of our
system. The image on the left is
the input and the two images on the right are the output of the system.
Auditory
Material Properties
We
have been developing a system to classify materials from the sound of impact.
In order to perform its goal the system extracts acoustic properties from the
sound and uses these values to determine the values of material properties,
such as internal friction. We have been working on the extraction of acoustic
properties to determine which are the most relevant and reliable properties to
be used in acoustic material classification tasks.
Research
Plan for the Next Six Months
Currently, the system
for obtaining shading and reflectance images from gray-scale images uses synthetic
images to learn the best filters for classifying the image derivatives. In the next six months, we will develop
means to train the classifier from real images. In addition, we will research how to isolate other intrinsic
characteristics of images.
We will continue
building tools for synthesizing sounds corresponding to standard shapes and
improve our system for recognizing materials based on their sound.