The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
654 
mind, that optimal stationing of the sensor is not always 
possible, especially in inner city areas, where traffic, property 
boundaries and other circumstances limit the choice for sensor 
stationing. 
In colour texture acquisition it has been shown, that automatic 
fusion of images acquired at multiple stations can minimize 
occlusions (Böhm, 2004). However, the key to this approach, i.e. 
redundant data acquisition, is only feasible since image 
acquisition with close-range cameras is fairly inexpensive. This 
approach is not recommended in terrestrial laser scanning, since 
each station is associated with considerable costs. The same is 
true for vehicle-based scanning systems, as either the number of 
scanners mounted on the vehicle or the number of drive-bys 
need to be increased. 
We therefore investigate the exploitation of self-redundancy for 
modelling facades from incomplete range data. This approach is 
based on the observation that typically façades of building are 
composed of repetitive patterns. This repetition of patterns can 
be used to substitute incomplete areas of the range data. One 
key question is how the structure of these repetitions can be 
detected and how it can be efficiently encoded. In this paper we 
propose a graph-based approach for the encoding of repetitions, 
which is efficiently derived from key point matching. Before we 
detail this procedure in section 5, we first give an overview of 
related work in section 2. In section 3 we give the fundamentals 
of our approach to façade modelling my LASERMAPs and in 
section 4 we show how defective areas can be substituted for. 
2. RELATED WORK 
In the computer vision literature the situation of replacing 
effective or occlude image parts has been dealt with extensively. 
Prominent solution include inpainting, a technique which is 
used to fill small gaps, typically by propagating linear structures 
from the border of the defective area into the area. A second 
approach is texture synthesis, which tries to copy repetitive 
texture to fill an occluded image area. There are many 
variations and also combinations of these methods, see for 
example (Criminisi, 2004). Our approach is most similar to 
exemplar-based inpainting methods. We are not aware of 
approaches specific to depth maps or range images. 
As mentioned earlier we have dealt with the case of occluded 
colour texture images (Böhm, 2004). As the acquisition of 
LiDAR data is inherently more time-consuming than image 
acquisition, it is unrealistic to assume a highly redundant multi 
station configuration to overcome occlusions, as it is possible in 
colour texture image synthesis. However, the idea to have a 
redundant description for one and the same area of an image 
still holds. And thus the proposed techniques for robust fusion 
can be transferred to LASERMAPS. 
Self-similarity, repetitive structures and symmetry of buildings 
and facades in particular have recently attracted great attention 
in the research field (Müller, Wonka et al. 2006; Ripperda and 
Brenner 2006). A number of successful applications were 
developed which exploit these properties (Müller, Zeng et al. 
2007). The aforementioned approaches use grammars to store 
the repetitive pattern of elements. This approach has shown to 
be successful for synthesizing complete façades and for creating 
variations of facades. In our work we chose a different 
representation scheme which is based on graphs. Our idea of 
using self-similarity for substituting incomplete data is also 
motivated by the work of (Pauly, Mitra et al. 2005). It differs in 
Figure 2. A prismatic building model with detailed roof 
structure and a registered point cloud acquired by ground- 
based LIDAR. 
that we select the examples used to fill the gaps from the dataset 
itself rather than using a database of models. 
The simple representation in the form of a LASERMAP enables 
us to use likewise simple image processing operators to extract 
structures. In order to detect repetitive patterns we use a similar 
procedure as has been proposed in (Wenzel, Drauschke et al. 
2007). It is based on feature point extractors typically used for 
registration of separate data sets. Matching these feature points 
within the same dataset detects self-similarity. 
3. FACADE MODELLING USING LASERMAPS 
In order to able to combine ground-based laser data with pre 
existing building models, the data has to be registered. There 
many possibilities to compute the registration, ranging from 
direct georeferencing, to manual and automatic alignment. Our 
approach to registration and georeferencing of terrestrial laser 
data and virtual city models is given in full detail in 
(Schuhmacher and Boehm, 2005). 
For the rest of this paper we assume that the registration has 
been computed and the range data is given with respect to the 
same coordinate frame as the building model. This initial 
situation of a point cloud registered to a building model is 
shown in Figure 1. The dataset depicts the president’s office at 
the Universität Stuttgart. The point cloud was acquired with a 
terrestrial laser scanner, a Leica HDS 3000, from more than 15 
stations. The data covers the facades of the building at a point 
density better than 20 mm. The large number of stations was 
necessary to minimize shadowing of occluding objects. By 
removing selected stations, we can now control the 
completeness or incompleteness of the data. 
Our method to modeling facades is motivated by concepts for 
modeling developed in computer graphics. In computer graphics 
the duality of coarse over-all geometry and fine detail has long 
been noted. The separation of the two is a fundamental 
modeling principle. Starting with the observations of Blinn 
(1978), that the effect of fine surface details on the perceived 
intensity is “primarily due to their effect on the direction of the 
surface normal ... rather than their effect on the position of the 
surface”, modeling concepts were developed, which keep fine 
surface detail separate as a perturbation of the normal direction 
or a displacement to the underlying coarser geometry.