HARDWARE-BASED TEXTURE EXTRACTION FOR BUILDING FACADES
Martin Kada
Institute for Photogrammetry (ifp), University of Stuttgart, Germany
Geschwister-Scholl-Strasse 24D, D-70174 Stuttgart
martin.kada@ifp.uni-stuttgart.de
Commission IV, WG 6
KEY WORDS: Extraction, Building, Texture, Visualization, Hardware, Graphics, Reconstruction
ABSTRACT:
The reconstruction of 3D city models has matured to the point where large data sets are now available. As most of the data collection
methods used are based on airborne sensors like e.g. aerial laser scanning or stereo imagery, the detailed geometry and material of
the building facades is typically not available. For visualisation purposes, however, the surface structure is essential to achieve a
good visual impression of the respective buildings. An efficient technique to provide the missing building facades is to extract
texture images from terrestrial photographs and map them to the polygonal faces of the reconstructed models. As the task of manual
texture extraction and placement is very time-consuming, an automatic approach is presented in this article that utilises the rendering
pipeline of modern 3D graphics cards. The problem of texture extraction can therefore be solved by using graphics algorithms that
are nowadays implemented in hardware and consequently are extremely fast. Since only parts of the building or even of a façade are
typically captured in one single image, self occlusions of the buildings are detected and several photographs taken from various
positions are fused to generate the final texture images. In order to gain high quality textures, the lens distortion of calibrated
cameras is corrected on-the-fly by the use of pixel shaders that are running on the programmable graphics processing unit.
1. INTRODUCTION 2). Such an approach is consequently not applicable for cap-
turing a large number of building façades.
The acquisition of 3D city models has been of major interest
for the past years and a number of algorithms are now avail-
able both for the automatic and semiautomatic collection of
3D building models. Based on measurement from aerial ste-
reo imagery or airborne laser scanner data, the geometry of
buildings can be reconstructed on a large scale. (Baltsavias,
Grün and van Gool, 2001) e.g. give a good overview of ex-
perimental systems and commercial software packages. One
major limitation of these approaches is, however, that the re-
sulting models have rather coarse façades. (Früh and Zakhor,
2003) present a method that merges ground based and air-
borne laser scans and images. The additional terrestrial data
naturally leads to more detailed façades.
Key market for this type of data is the visualisation in the
context of city planning, three-dimensional car navigation,
virtual tourism information systems and location based ser-
vices. In addition to pre-rendered movies of the virtual envi-
ronments where the user has no freedom of movement, real-
time visualisation is getting more and more important. (Kada Figure 1. A 3D landscape model of Stuttgart rendered in a
et al., 2003) show e.g. that literally a complete city can be in- real-time visualisation environment. All facade
teractively displayed in 3D on today's consumer PC systems textures of buildings located in the main
(see Figure 1). pedestrian area were manually captured.
For the photo-realistic visualisation of urban landscapes, the
material of the building façades is essential for the visual im-
pression. An efficient technique to model building façades is In this article, an approach is described that automatically ex-
to extract texture images and place them on the coarse, po- tracts façade textures from terrestrial photographs and maps
lygonal faces of the reconstructed models. Whereas roof im- them on geo-referenced 3D building models. If the exterior
ages can easily be acquired from aerial photographs, such an orientation of the camera is known, a transformation can be
approach is not feasible for the building facades. It is there- computed that projects the polygonal faces of the building
fore inevitable to use terrestrial images as the source for model into the image. (Klinec and Fritsch, 2003) determine
high-quality facade textures. The manual texture extraction the rotation and translation of the exterior orientation by
and placement is, however, a tedious task and can easily take searching for correspondences between object and image fea-
up to several days per building for good results (see Figure tures and use them in a photogrammetric spatial resection.
420
Internat.
Figure 2
The pre
the exte
scribes
3D obje
model t
to visuz
spondin
eg.aV
approac
tions cz
distortic
By usi
whole p
tive too
graphs .
texturec
the mox
can be |
tortions
ware so
Because
functio1
can be ı
2. H
The te:
utilises
commo
develop
discuss
21 Gi
The ext
involve
pixel d.
increase
exploit
is gene
units ((
graphic
process
functio:
utilised
