f real world 
nages of the 
| in order to 
iges undergo 
ybject can be 
ct. The shell 
ne the model 
yrmal vector 
e orientation 
carved away 
ject. 
avities on an 
Ve refine the 
by our shell 
our proposed 
z and Dyer, 
gorithms use 
om the other 
e points in a 
in the input 
xel coloring 
he visibility 
f pooling the 
, we perform 
ich distortion 
is described. 
lained. The 
m silhouette 
4 introduces 
ster 5, image 
| algorithm is 
e are using à 
» acquire still 
> the interior 
acquisition is 
und. Multiple 
in a circular 
2.1 Camera Calibration and Image Orientation 
Prior the image acquisition, the camera should be calibrated. 
We calibrated the sensor using several images with a calibration 
object having three perpendicular square planes and 25 control 
points on each side. Since we use an off-shelf CCD camera, we 
switch off the auto-focus, so that the focal length remains fixed 
throughout the whole process. 
In a second step, the object is placed inside the calibration 
frame in order to define some natural control points accurately. 
We performed a bundle block adjustment with all the images, 
which delivered the interior camera parameters as well as the 
coordinates of the control points, which were initially 
introduced as new points. 
In order to compute the object’s model, the images should be 
oriented, the rotations and the position of the cameras should be 
known. In many cases we cannot mark control points on the 
objects, therefore natural textures can be used. 
The images were adjusted in a bundle block adjustment process. 
We used enough tie points in all images in the circular camera 
setup to perform a bundle block adjustment, covering all 
images. We achieved very accurate results for the image 
orientations, using the previously calibrated camera. The image 
projection centers had accuracies of 1-2 mm, the rotation were 
determined with 0.05-0.1 gon. 
3. APPROXIMATE MODEL 
One of the well-known approaches to acquire 3D models of 
objects is voxel-based visual hull reconstruction, which 
recovers the shape of the objects from their contours. 
A silhouette image is a binary image and easily obtained by 
image segmentation algorithms. Image pixels indicate if they 
represent an object point or background point. Since the blue 
background that we use is sufficiently homogeneous, we can 
easily define a hue domain, which is considered background. A 
pixel's position in the IHS-colorspace is examined in order to 
decide if it represents background or object. We performed the 
image segmentation using the academic software HsbVis. 
  
  
Figure 1: Intersection of silhouette cones 
To start with, we define a 3D discrete space which contains 
only opaque voxels with the value *255" representing object 
points. In order to compute the silhouette cone, we projected all 
the cube's voxels into every image. If the image coordinate 
defines a background pixel; the voxel is labeled transparent by 
giving it the value *0" which means the voxel of interest now 
represents empty regions in the voxel cube. The volume 
intersection algorithm intersects all silhouette cones from 
multiple images to achieve the estimate geometry of the object, 
which is called the object’s visual hull. See (Kuzu and 
Rodehorst, 2000) for more details. 
In Figure 1 you see the intersection of the silhouette cones 
acquired, using 1, 3, 5 and 9 images. As you will notice with 
increasing number of images this method obtains better 
approximations to the objects true shape. 
As shown in Figure 2, concavities cannot be recovered with this 
method since the viewing region doesn't completely surround 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
the object. The accuracy of the visual hull depends on the 
number of the images and the complexity of the object. 
E 
— et 
ce po d 
aae 
Figure 2: Concave areas in visual hull reconstruction 
However, since the result encloses the largest possible volume 
where the true shape lies and the implementation is 
straightforward and easy, it is an attractive. method for 
applications where the approximate shape is required. We use 
the visual hull as the first step of our reconstruction algorithm 
and we consider the shell carving algorithm as a refinement 
method to carve away the necessary voxels in the concave areas 
of the visual hull for a more precise reconstruction. 
4. COMPUTATION OF VISIBILITY INFORMATION 
It is crucial to find out which voxel is visible in which image. 
We will use a line tracing algorithm to check each voxel along 
the line, whether it is background or object voxel. As soon as an 
opaque voxel is encountered, the initial voxel can be considered 
occluded. When the line exits the defined voxel cube, it can be 
stopped, assuming that the voxel is visible. Whether lying on 
the backside or occluded by another voxel, the algorithm will 
correctly tell if the voxel is visible or not. 
   
Figure 3: Considering occluded areas 
In Figure 3, we show why the knowledge of visibility can be 
crucial. If we take a closer look at the vase, we will see that the 
handle is occluding some voxels for some specific images. 
Hence these images, which theoretically have the best view to 
the occluded voxels, concerning the viewing angle, cannot see 
the voxels and therefore should not be considered. From the set 
of remaining images, the best candidate needs to be chosen.