482 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
Fig. 3: Extracted surfaces of the sequence shown in figure 1 and 
2a) with extracted textures. The small dots show the estimated 
camera positions. 
It can be seen that the textures on the surfaces are looking tilted 
in movement direction. This is caused by the angular view 
including a little upwards looking angle that is not correctly 
estimated and leads in fact to a tilt of the surfaces. The textures 
are then extracted correctly from the image sequence and pasted 
on the tilted surfaces. This error is proportional to the up 
looking angle of the camera. Another interesting fact is the 
perpendicular façade (Fig. 3b). Although a tree is standing 
before the façade, the algorithm could estimate a surface. In this 
case, even two surfaces were detected. The right part of the 
façade shows two non occluded windows and was estimated as 
façade. The position error is caused by the nadir view of the 
camera to the façade causing only rough positioning 
information. The left part was detected at the position of the 
tree. Most of the texture of this part shows the trunk and 
branches of the tree. Between the second (Fig 3c) and the third 
(Fig 3d) big façade at the right, no façade can be seen. In the 
original image sequence (Fig. 1 and 2a) one can see a tree and 
in addition some flag staffs. This produces too many wrong 
pairs of points of interest to detect the surface plane. 
5.2 Comparison with the given 3d model and measured 
camera path 
The façades are not extracted in their correct size. In figure 3, at 
the left comer the façade textures shows only two windows left 
of the bright entrance gallery. In fact, there must be three. The 
and not at the border. Only the right border of the façade is 
found correctly because of the high number of points of interest. 
These points are extracted from the strong image border 
between the façade and the tree. In the up direction, the façades 
do not contain a tilt. But this is clear, because the vertical 
direction is set fix in the algorithm and so the estimated planes 
are forced to be almost vertical. Figure 4 shows a front view of 
the building with both the measured and estimated model. 
The image sequence starts at the left. The planes are shifted 
with about 3 meters to the right. The size of the left façade 
plane is almost the correct size of given façade except for the 
tilt in the moving direction of the camera (see Fig. 3). The 
central façade is also detected with almost the correct size, but 
shifted. The small rectangular plane at the left edge of the 
central façade is the wrong detected surface belonging to the 
tree (see Fig. 3). For the right façade, the height is estimated 
almost correctly but the plane is stretched in moving direction. 
This is caused because the perpendicular small façade (Fig. 3 at 
the right) is not detected and so some points of this small façade 
are added to the front façade. Except this, this façade shows 
only the shift. This shift is caused by the errors of the given 
GPS position of the camera. This error can be seen in figure 5. 
/ 
/ 
- - - - r - ■-* ■. : ' WV • * 
rrrrrrr'r V W ,W w V W V w 
Fig. 5: Top down view of a building façade with measured GPS 
camera path (yellow) and estimated camera path 
(green). The estimated surface plane is red. 
It can be seen that the measured camera path from the GPS is 
leading towards the façade. In fact, the camera path must be 
parallel to the façade, so this is an error in the GPS positioning. 
When the estimated camera path is moved to the GPS path, 
there remains the shift error as seen in figure 4 and this angular 
error. When including the estimated surface planes in the 
positioning, the results are looking like in figure 5. The 
Fig. 4: Front view of the building. Dark blue and green: Measured model, Light blue: estimated façade planes, Yellow: 
measured camera path, Light green: estimated camera path 
façade does not contain its roof and floor borders. This is on the 
one hand caused by the tilt; on the other hand, the surface is 
estimated as to small. This is caused by the point extraction 
density, which leaves a small . band around the façade 
unexplored because the points of interest are found at the façade 
estimated camera path is almost parallel to the façade and the 
estimated surface plane is almost parallel to the model façade. 
Another aspect is the distance between the camera and the 
building which is estimated very good with an error of about 
half a meter for this scenario. The angular error of the GPS