CIP A 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
107 
Figure 6. First and last image of an image sequence taken with the Sony camcorder (left below) and two views of the VRML 
model fully automatically generated from the sequence (top and right). 
Figure 6 shows the first and the last image of a sequence taken 
with the Sony camcorder and two views of the resulting 
VRML model. Of the 1500 images we have taken only seven 
to minimize the computational effort. Overall, we obtained 71 
7-fold, 60 6-fold, 55 5-fold 24 4-fold, and 4 3-fold points after 
robust adjustment and an error of 0.06 pixels. For this camera 
only the range of the focal length is known, but not the pixel 
size or the metric size of the sensor, i.e., the parameters of the 
camera matrix are unknown. Our auto-calibration scheme 
resulted into 1.66 for the ratio of principal distance to image 
width and 2.38 for the ratio of principal distance to image 
height. The precision in image space was 0.25 pixels. It should 
be noted that the results have been obtained without user 
interaction and with the same set of parameters as for the 
above sequence. The two views of the VRML model generated 
for the points and the cameras shows that the metric geometry 
of the scene has been recovered fairly well. 
4.3 Structure computation 
Our approach for the computation of 3D structure (Mayer, 
2003), i.e., a disparity image in the first place, from an image 
pair or triplet is based on the algorithm for cooperative 
disparity estimation proposed in Zitnick and Kanade (2000). 
The respective image pair is resampled along the epipolar 
lines. Matching scores are computed by cross correlation and 
absolute differences and written in a 3D array made up of 
image width, height, and disparity range. As the computational 
effort depends directly on the range of the disparity, we 
compute this range by projecting the points reliably 
determined for the image triplet onto the epipolar lines. 
Figure 7: Disparity map (left) and visualizations based on the disparity map