The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
3d point cloud from the image sequence three different cases of 
façades have to be investigated. 
1. The first façade type is standing parallel to the street 
and the camera is moving along this façade. These façades 
fulfil the conditions for the 3d point extraction, because the 
points are moving through the image from the right to the 
left and so their relative 3d position can be determined 
from their movement. 
2. The second class on façades is the class of occluded 
invisible façades. They, of course remain invisible as holes 
in the generated point cloud. 
3. The third class of façades is standing perpendicular to 
the street. Those façades are not moving along the camera, 
but are changing their scale as they are moving towards the 
camera. For those façades, the movement of points is very 
low and thus it is very difficult to estimate the correct 
coordinates. 
4. 
In addition to the limitations in the viewing position of the 
camera the low resolution and low grade of details cause a 
small number of points of interest. 
3.2 Description of the quality measurements 
It is necessary to make a quality investigation including the 
estimated camera path, the position and completeness of 
façades and the extracted textures of the façades in comparison 
to the given 3d building model of the GIS database and the 
recorded GPS camera path. For all images and GPS positions a 
synchronized time-code is stored. The camera path of the GPS 
is corrected using a Kalmann filter and interpolated for every 
time-code corresponding to an image of the sequence. Now, the 
first and the last position of the estimated camera from the 5- 
point pose estimation are moved to the interpolated positions of 
the GPS path with the corresponding time codes. This leads to a 
scale, rotation and translation of the estimated model onto the 
given 3d model and the GPS path. 
The estimated planes are now transformed to the world 
coordinate system. The scale factor is calculating comparing the 
length of the given camera path and the length of the estimated 
camera path for the first and last camera position. Afterwards, a 
3d rotation and translation is calculated to rotated and move the 
estimated camera path onto the given measured one. Because of 
the assumption that the GPS path is not afflicted with an error, 
the corresponding surfaces of the given building model and the 
generated planes are given by the smallest error in their 
orientation and the smallest translation vector of their 
barycenters. To avoid a systematic error caused by the GPS 
position, the translation vectors of the barycenters of all 
generated planes to their corresponding model surface are used 
to calculate a mean translation vector. The remaining 
translation vectors of the planes to the surfaces are the 
remaining positioning errors of the planes, the generated mean 
translation vector is used to move the generated camera path of 
the image sequence. The distance between this path an the GPS 
path can be caused by the 5-point algorithm, an incorrect 
camera calibration or the inaccuracy of the GPS positions. 
In addition to the positioning error of the planes, the 
completeness is a criterion of the quality of the surface 
reconstruction. For every estimated plane its length and height 
are determined calculating its bounding rectangle and compared 
to the length and height of the given corresponding façade. 
Caused by the reconstruction from points of interest, the 
generated planes are estimated to be a little bit smaller than the 
original façade. 
4. EXPERIMENTS 
The camera that was used for the acquisition of the test 
sequences offers an optical resolution auf 320x240 (FLIR 
SC3000) pixel with a field of view (FOV) of only 20°. The 
SC3000 is recording in the thermal infrared (8-12 pm). On the 
top of a van, the camera was mounted on a platform which can 
be rotated and shifted. Different scenarios of image sequences 
were acquired. The first scenario (Figure 2a) deals with several 
small façades belonging to one building block. The second 
(Figure 2b) shows a long façade with regular structure and a 
specific entry with overlap. The third scene (Figure 2c) shows a 
long façade with different structures. The forth scenario (Figure 
2d) deals with bridges between buildings crossing a street. 
Fig. 2: a) several façades with occlusion, b) façade with regular 
structure, c) façade with irregular structure, d) 
façade with building bridge 
For scenario 2 and 3, the situation is quite different. Both 
scenarios consist of only one long façade. This façade can be 
estimated quite easy because there are no occlusions of the 
façade and the complete façade is almost in one plane. Scenario 
4 has to deal with two building bridges crossing the street. 
Those bridges are separating the façade into segments. The 
bridges cannot be detected. 
5. RESULTS 
5.1 Extracted surface planes and camera path 
In general, the surface estimation works well for façades going 
parallel to the street. Façades standing perpendicular to the 
street are much more difficult to extract correctly. For the scene 
in figure 2a, three parallel surfaces can be extracted. For these 
surfaces, the relative camera path containing relative camera 
positions for all images of the sequence is generated. The result 
of the automated extraction is shown in figure 3.