International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 
a scaling part Tp = Ts : T,,. Since the target shape is a rectan- 
gle, the image can be transformed onto a unit square. This will 
lead to an orthographic projection, however the width and height 
of the image are incorrect. The correct scaling will be applied to 
the cropped image at rendering stage, when the image is mapped 
onto the corresponding facade. With this simplification only the 
four corner points of the facade have to be identified in the image 
and no control information is required. The advantage of such 
an approach is that it does not require the determination of the 
exterior orientation and neither does it require the calibration of 
the camera. Furthermore no control points in object space have to 
be determined. The disadvantage of using only a single image is 
that any object occluding the facade will be mapped as well and 
thereby disturbing the texture image. To avoid these occlusions 
manual stitching of images is required. Despite these disadvan- 
tages the approach has shown to be quite successful and several 
hundred buildings have been textured using several thousand ter- 
restrial images (Kada et al., 2003). 
Coorg and Teller (1999) have introduced an automated approach 
for reconstructing vertical façades from a set of images using a 
space sweep approach. The method relies on controlled imagery, 
ie. in photogrammetric terms the exterior orientation of each 
image is assumed to be known. Textures are computed from a set 
of images by a weighted median estimation process. 
Wang et al. (2002) have contributed to this work with a further 
improved method of computing texture. They account for three 
sources of influence on the façade image: the occlusion by mod- 
eled objects, the obliqueness of a certain portion of the façade 
with respect to the camera station and the occlusion by unmod- 
eled objects. The texture image is computed in an iterative man- 
ner using a weighted-average approach. In addition to the texture 
they present a method to compute the 3D façade relief as well. 
Bornik et al. (2001) presented a photogrammetric approach for 
deriving texture maps. The geometry of buildings is reconstructed 
by means of photogrammetry using multiple views. Camera cal- 
ibration data and exterior orientation from the photogrammetric 
step are then used to compute texture images. The texture image 
is synthesized from several views again using a median filter. 
For our work we aim at keeping the simplicity of an approach 
purely based on perspective transformation. Yet we want to elim- 
inate the need for manual stitching to suppress occlusions and 
thus enhance the approach to use multiple views. The main idea 
as mentioned. above is to create pixel-level registered image se- 
quences from these views. Regarding multiple images as a se- 
quence allows us to apply background estimation techniques to 
eliminate occlusions. 
The next section gives a quick overview of background estima- 
tion and introduces simple examples where several images from 
a single photo station are used, allowing for the elimination of 
moving objects. 
3 BACKGROUND ESTIMATION 
The process of background estimation or background mainte- 
nance is frequently encountered in video surveillance systems, 
where a fixed camera is directed into a hallway, onto a building or 
to a road crossing, and so on. The sequence of images obtained 
from the camera is compared to an existing background image. 
Using simple background subtraction changes in the scene can be 
observed. These changes can be caused by people, cars or other 
objects, which are to be identified, tracked or otherwise detected. 
The difficulty in such a system lies in obtaining the correct back- 
ground image. This is especially true when no image of the back- 
ground free of occluding object is available. A second reason for 
868 
. Vol XXXV, Part BS. Istanbul 2004 
  
  
Figure 1: Façade textures can compensate for lack of geometric 
detail. The improvement in visual appearance can be seen from 
the above two images, where the second image is based on the 
exact same geometry and only texture was added. 
  
Figure 2: The model of a single building. All façades are approx 
imated by planar surfaces and bounded by rectangular curves. 
the 
ba 
otl 
tio 
lis 
len 
fra 
ing 
ser 
In: 
qu 
tec 
Col 
of 
col 
sta 
prc 
res 
Th 
the 
abl 
eff 
cia 
the 
the 
pre 
Sin 
eff 
is ı 
of «