| to Edge 
alysis and 
ensor Data 
ternational 
ng, XXXII 
M. 2001. 
action in 
mmetry &. 
nition and 
National 
ation. US 
LO. 1997 
ternational 
ig, Vol. 32, 
C., Ebner, 
ban Areas. 
"om Aerial 
A. Gruen, 
Lisse, pp. 
… Schmidt, 
ved Urban 
BellSouth 
echnology, 
1g, 60:339- 
x Detection 
of ISPRS 
: Meeting 
S France. 
Effectively 
5 Images. 
; 2003. 
anbul 2004 
  
DETAILED RELIEF MODELING OF BUILDING FACADES FROM VIDEO SEQUENCES 
Wolfgang v. Hansen^, Ulrich Thónnessen^, Uwe Stilla^ 
*FGAN-FOM Research Institute of Optronics and Pattern Recognition 
Gutleuthausstr. 1, 76275 Ettlingen, Germany s 
wvhansen Q fom.fgan.de 
"Photogrammetry and Remote Sensing, Technische Universitaet Muenchen, 
Arcisstr. 21, 80280 Muenchen, Germany 
Working Group III/2 
KEY WORDS: Building, Surface, Reconstruction, Photo-realism, Texture, Urban, Close Range, Video 
ABSTRACT 
Three dimensional building models have become important during the past years for various applications like urban 
planning, enhanced navigation or visualization of touristic or historic objects. Altough for some applications geometric 
data alone is sufficient, for visualization purposes a more realistic representation with textured surfaces is necessary. The 
associated textures from buildings are extracted either from airborne imagery or, especially for facades, from images taken 
. by ground based cameras. For very high demands on photoreal 
istic quality, textures mapped on simple geometric models 
like polyhedra or regular surfaces are not sufficient because relief structure is not preserved. This leads to an unrealistic 
and artificial impression on close-up views. Relief structures are beneficial for large scale models that are closely inspected 
within a limited area in a virtual world. In this paper the extraction of reliefs to improve existing planar surfaces of wire 
frame models of buildings is described. Given several uncalibrated views onto a surface of the polyhedral model, a depth 
map is estimated by correlation. The underlying plane is used to guide the correlation in order to detect outliers and to fill 
in homogeneous areas. 
1 INTRODUCTION 
Three dimensional building models have become impor- 
tant during the past years for various applications like 
urban planning, enhanced navigation or visualization of 
touristic or historic objects (Brenner et al., 2001). Build- 
ing models are typically acquired by a (semi-) automatic 
processing of laser scanner elevation data or aerial imagery 
(Baillard et al., 1999). Although for some applications geo- 
metric data alone is sufficient, for visualization purposes a 
more realistic representation with textured surfaces is nec- 
essary. 
The associated textures from buildings are extracted ei- 
ther from airborne imagery or, especially for facades, from 
images taken by ground based cameras (Teller, 1998). 
For very high demands on photorealistic quality, textures 
mapped on simple geometric models like polyhedra or reg- 
ular surfaces are not sufficient because relief structure is 
not preserved. This leads to an unrealistic and artificial im- 
pression on close-up views. Relief structures are beneficial 
for large scale models that are closely inspected within a 
limited area in a virtual world. For architectural and touris- 
    
E: 
Coarse Model 
  
Figure 1: Refinement of a coarse model by images. 
tical applications it is advantageous to dispose of enhanced 
facade models with relief information in order to improve 
visualization of door and window openings as well as or- 
naments. 
Facade relief information can be extracted by measure- 
ment of depth information using à terrestrial laser scan- 
ner or by photogrammetric analysis of the same images 
also acquired for texture mapping. Nowadays, video cam- 
eras are widely available as an inexpensive source of data. 
Data acquisition often is not planned as thoroughly as for 
a true photogrammetric campaign and therefore provides 
more challenging data. For some historic buildings that 
have been destroyed or otherwise changed, images taken 
by tourists or non-professionals might be the only source 
of information available (Grün et al., 2003). Because the 
resolution of images of a video camera is low compared to 
that of a photographic camera one of the important ques- 
tions is the amount of geometric accuracy and level of de- 
tail that can be expected from such acquisitions. 
In this paper the extraction of reliefs to improve existing 
planar surfaces of wire frame models of buildings is de- 
scribed. Given several uncalibrated views onto a surface of 
the polyhedral model, a depth map is estimated by correla- 
tion. The underlying plane is used to guide the correlation 
in order to detect outliers and to fill in homogeneous areas. 
Basically, two different methods to reconstruct 3D objects 
from images for visualization in a virtual environment can 
be identified. One is to generate surface models directly 
by dense stereo matching (Pollefeys, 1999). No explicit 
knowledge of the exact shape of the objects is required 
which makes the approach simple in the sense that only