CIP A 2003 XIX th International Symposium, 30 September-04 October, 2003, Antalya, Turkey 
105 
e terrain 
:omplex. 
a small 
ould be 
from an 
dent ray 
; point 
:essfully 
from the 
le Rollei 
a Sony 
cations). 
res of a 
a rigid 
mera. In 
illy, i.e., 
settings 
ley were 
ic 5 used 
il length 
i for the 
f, was 
roviding 
e size as 
> of the 
shorter 
ct some 
resented 
3. PHOTOGRAMMETRIC RESTITUTION 
A subset of 10 Rollei images covering a part of the first 
courtyard was selected to compare the 3-D reconstruction 
process normally used in close-range photogrammetry with 
methods preferably applied in computer vision. The digital 
images were taken parallel to the object facade having 
relatively short distances between the camera positions. Image 
data were obtained by manual pointwise measurement within 
the AICON DPA-Pro and PhotoModeler software. Then, 
interior and exterior orientation parameters as well as 3-D 
coordinates of object points were determined by self 
calibrating bundle adjustment. Fig. 3 shows the result of the 
visualization of this part of the castle. 
Figure 3. Wartburg Castle: Part of the first courtyard 
4. FULLY AUTOMATIC SEQUENCE ORIENTATION, 
AUTO-CALIBRATION, AND 3D RECONSTRUCTION 
The goal of this part was to investigate, whether a sequence of 
images, for which the only thing known is, that they are 
perspective and that they mutually overlap, is enough for a 
metric reconstruction of the scene. Additionally we were 
interested to compare the automatically computed camera 
calibration information to given calibration information. 
4.1 Sequence Orientation Based on the Trifocal Tensor 
While other approaches use image pairs as their basic building 
block (Pollefeys, 2002), our solution for the fully automatic 
orientation of an image sequence relies on triplets which are 
linked together (Hao and Mayer, 2003). To deal with the 
complexity of larger images, image pyramids are employed. By 
using the whole image as search-space, the approach works 
without parameter adjustment for a large number of different 
types of scenes. 
The basic problem for the fully automatic computation of the 
orientation of images of an image sequence is the 
determination of (correct) correspondences. We tackle this 
problem by using point features and by sorting out valid 
correspondences employing the redundancy in image triplets. 
Particularly, we make use of the trifocal tensor (Hartley and 
Zisserman, 2000) and RANSAC (random sample consensus; 
Fischler and Bolles, 1981). Like the fundamental matrix for 
image pairs, the trifocal tensor comprises a linear means for 
the description of the relation of three perspective images. 
Only by the linearity it becomes feasible to obtain a solution 
when no approximate values are given. RANSAC, on the other 
hand, gives means to find a solution when many blunders 
exist. 
Practically, first points are extracted with the Forstner 
operator. In the first image the number of points is reduced by 
regional non-maximum suppression. The points are then 
matched by (normalized) cross-correlation and sub-pixel 
precise coordinates are obtained by least squares matching. To 
cope with the computational complexity of larger images, we 
employ image pyramids. On the coarsest level of the image 
pyramid, with a size of approximately 100 x 100 pixels, we use 
the whole image size as search space and determine 
fundamental matrices for image pairs. From the fundamental 
matrices, epipolar lines are computed. They reduce the search 
space on the next level. There, the trifocal tensor is 
determined. With it a point given in two images can be 
projected into a third image, allowing to check a triple of 
matches, i.e., to sort out blunders. For large images, the 
trifocal tensor is also computed for the third coarsest level. To 
achieve highly precise and reliable results, after the linear 
solution projection matrices are determined and with them a 
robust bundle adjustment is computed for the pairs as well as 
for the triplets. 
To orient the whole sequence, the triplets are linked. This is 
done in two steps. First, the image points in the second and 
third image of the nth triplet are projected into the third image 
of the n plus first triplet by the known trifocal tensor for the n 
plus first triplet. As the (projective) 3D coordinates of the nth 
triplet are known, the orientation of the third image in the 
projective space of the nth triplet can be computed via inverse 
projection. To obtain high precision, a robust bundle 
adjustment is employed. In the second step. 3D coordinates in 
the coordinate system defined by the nth triplet are determined 
linearly for all points in the n plus first triplet that have not 
been computed before. The solution is again improved by 
robust bundle adjustment. Starting with the first image, this 
incrementally results into the projective projection matrices for 
all images as well as in 3D points. After having basically 
oriented the sequence on the two or three coarsest levels of the 
image pyramid, finally, the 3D points are projected into all 
images via the computed projection matrices. The resulting 
points are then tracked over one or two levels through the 
pyramid. 
Figure 4 gives results for the first four images of the sequence 
taken with the Rollei d30 metric 5 camera. One can see that the 
points have been tracked pretty well even for the wall close to 
the camera, where the disparities are rather large. For the 
whole sequence of ten images we have obtained 56 10-fold, 41 
9-fold, 91 8-fold, 71 7-fold, 55 6-fold, 30 5-fold, 41 4-fold, 
and 44 3-fold matches after robust adjustment with a standard 
deviation of 0.08 pixels.