object cannot be recovered since the viewing region doesn’t
completely surround the object. The recent attempts are based
on voxel coloring algorithms (Seitz and Dyer, 1997),
(Culbertson et al, 1999), (Kuzu and Sinram, 2002). These
algorithms use color consistency to distinguish surface points
from the other points in the scene making use of the fact that
surface points in a scene project into consistent (similar) colors
in the input images. In this paper we are using a color image
matching algorithm to refine the visual hull of the object.
Prior to the object reconstruction, the camera was calibrated, as
it is described in chapter 2.1. Some control points on the object
itself were defined in a local coordinate system, which were
used to perform a bundle block adjustment with all 22 images.
In chapter 2.2 the orientation process is explained.
The reconstruction of the model is described in chapter 3, using
shape from silhouette and color image matching techniques.
2. SYSTEM CONFIGURATION
The system we use consists of a simple CCD video-camera with
the ability to acquire still images. Furthermore, we need a
calibration object to compute the interior orientation parameters
of the camera. The object is placed in front of a homogeneous
blue background to distinguish background from object pixels.
The object is then rotated, resulting in a circular camera setup.
In the following chapters we would like to state the camera
calibration briefly, the introduction of control points on the
object and finally the bundle block adjustment.
2. Camera Calibration
As a very basic precondition to any subsequent spatial object
reconstruction, the sensor has to be calibrated. In our case, there
is no way of using a calibration certificate, since we are using a
standard video camera with auto-focus. Although the focus can
be fixed, we cannot assure that has been unchanged since the
last use.
So we calibrated it anew, using several images with a special
calibration object, as shown in figure 1.
Figure 1. Calibration object with 75 spatially well distributed
control points.
The camera parameters were calibrated in a bundle block
adjustment with self calibration, using five images and 75
control points each. The system had 750 observations and 33
unknowns, 6 for each image and three for the camera.
Additional parameters were intentionally ignored, since
previous calibrations have shown that they are neglectable.
The resulting parameters for the camera were as follows:
Calibrated focal length: c = 34.133 + 0.185 mm
Principal point: xp = 0.211 + 0.146 mm
Yp = 0.050 + 0.137 mm
As mentioned, the focus remained fixed throughout the
subsequent processes.
2.2 Image orientation
Prior to the image orientation, we had to apply some control
points to the object itself. It was out of question to mark points
artificially, so we had to choose ‘natural’ textures, instead.
We used the coordinates of the calibration object to define a
local coordinate system, in which we derived the control points
on the object. This is an arbitrary system, without relation to a
geodetic reference system. Figure 2 illustrates the control points
on the object, which served as reference for the subsequent
image orientation.
Figure 2. Some unique object control points.
For an accurate object reconstruction the exact image
orientation must be known. Consequently, the images were
adjusted in a bundle block adjustment, using 22 images in a
circular setup. Figure 3 shows a visualization of the setup
situation.
Figure 3. The virtual camera setup, using VRML-visualization.
As depicted in figure 2, we were able to use control points on
the front part of the object. Using enough tie points in all the
images, it was possible to perform a bundle block adjustment
with all the surrounding images. With the previously calibrated
camera, we managed to achieve very accurate results. The
image projection centres had accuracies of 1-2 mm, the rotation
were determined with 0.05-0.1 gon.
—170—
3.2
be
evi
pp
co
pr
sil
ge
nu
the
thi
