CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
orientation and object reconstruction. The object reconstruction 
is not discussed in this paper because it is presented in 
(Hrabacek and Heuvel, 2000). For camera calibration five 
images were selected. Figure 1 gives an overview of the images 
used in the experiments. 
1.3. Structure of the paper 
The paper is split in two main parts. First the camera calibration 
is discussed in section 2, and then the relative orientation of the 
four corner images is presented in section 3. These two sections 
start with a subsection in which the approach is briefly 
explained, followed by the results of the experiments using the 
images of the CIPA reference data set. Conclusions are 
presented in section 4. 
2. CAMERA CALIBRATION 
2.1. The method for camera calibration 
The method for camera calibration is summarised in the 
following three steps (Heuvel, 1999a): 
1. Extraction of straight image lines 
2. Detection of the object orientation of the image lines 
3. Estimation of camera parameters from parallelism and 
perpendicularity constraints on image lines 
The first two steps can be performed manual as well as 
automated. In the latter case, a line-growing algorithm is used 
for image line extraction (Heuvel, 2001), and a vanishing point 
detection method is applied for the detection of the three 
dominant object orientations (Heuvel, 1998). The quality of the 
estimated parameters is dependent on a correct vanishing point 
labelling of image lines performed in step 2. With the camera 
parameters unknown, the automatic detection of the three 
vanishing points that correspond with edges of the three 
orthogonal object space orientations is more critical than when 
using a calibrated camera. Two factors play a role. First, 
unknown lens distortion hinders the straight line detection and 
prohibits the detected lines intersecting in one point in image 
space. Second, unknown focal length and principle point make it 
impossible to limit the search space after detection of one or two 
vanishing points. As a result each vanishing point is detected 
independent of previously detected vanishing points. The 
procedure below is applied to five images of the CIPA data set. 
The lens distortion is determined first, followed by the other 
parameters of the camera model, i.e. focal length and principle 
point. 
1. Start with vanishing point detection for those images that 
contain only one façade of the building. For these images 
only two vanishing points are to be detected, one for the 
vertical object orientation, and one for the horizontal 
object edges. 
2. If only images with two (presumably orthogonal) façades 
are available, only the vanishing point that corresponds to 
the vertical object orientation is detected and its lines used 
for estimation of the lens distortion. This approach 
assumes limited camera tilt and rotation around the optical 
axis. 
3. Estimate lens distortion using the detected and labelled 
lines of at least one, but preferably more images. In the 
next steps the estimated lens distortion is eliminated from 
the observations. 
4. Detection of three vanishing points. Three-point 
perspective imagery (for example image 10 in Figure 1) is 
required for the estimation of the focal length and the 
principle point. When the optical axis is nearly horizontal 
and thus a one-point or two-point perspective remains (see 
the images in the bottom row of Figure 1) the principle 
point location in horizontal direction (camera x-axis) 
cannot be estimated, or only with very low precision. 
5. Estimate focal length and principle point using the 
detected and labelled lines of at least one but preferably 
more images. 
2.2. Camera calibration using the CIPA data set 
Image lines are extracted for all the selected images of the CIPA 
data set. For the line-growing algorithm used for straight line 
extraction two parameters were set. First the parameter for the 
minimum gradient strength was selected in such a way that most 
of the characteristic features of the building - especially the 
windows - were extracted. The second parameter is the 
minimum length in pixels of an extracted image line. This 
parameter was fixed to 30 pixels for all the images used for 
camera calibration. 
Lens distortion 
Detection of two vanishing points was performed on two images 
(numbers 8 and 9) that show only one face of the building. This 
is step 1 of the procedure described in the previous section. The 
result is shown in Figure 2. In this and following figures the 
image lines are colour-coded using the line labelling results of 
the vanishing point detection. 
Figure 2 : color-coded image lines of the vanishing point 
detection for image 8 (top) and 9 (bottom). 
For the estimation of the lens distortion (parameter kl) 
parallelism condition equations for the lines in images 8 and 9 
have been used. The estimated value for kl is -0.570 xlO' 3 . This 
value is 7.0 times its estimated standard deviation and thus