nbul 2004 
  
>xtured 
s the first 
level, the 
ces. From 
d a higher 
ompose a 
story into 
or the end 
ontage is 
stitute the 
nd can be 
ge plane. 
n level is 
5 and its 
tage faces 
vent from 
. modeled 
5 to 3D 
hem into 
per three 
topologic 
whatever 
= division 
ribe and 
xtures for 
rithms for 
complex 
sie pour 
intique en 
formation 
2 pages. 
outils de 
S11. 3D 
e. 
sition and 
ORS "Vol. 
for three 
e on GIS 
rlands. 
  
3-D BUILDING RECONSTRUCTION FROM UNSTRUCTURED DISTINCT POINTS 
Chiung-Shiuan Fu 
Jie Shan 
Geomatics Engineering, School of Civil Engineering, Purdue University 
550 Stadium Mall Drive, West Lafayette, IN 47907-2051, USA 
Commission III, WG III/7 
KEY WORDS: three-dimensional, point cloud, reconstruction, modeling, visualization 
ABSTRACT: 
This paper presents a model-based approach for 3-D building reconstruction unstructured distinct points. The data structure of a 
building is an unstructured roof point cloud digitized by operators. Constructive Solid Geometry (CSG) model is applied in our 
approach for a complex roof which can be decomposed into three types of primitive models. The primitive model is represented as a 
tree. The leaves contain primitives and the nodes store Boolean set operators to combine primitives to form a whole building. There 
are three steps in this approach. The key process during the entire reconstructing procedures is rectangulation, which is to form points 
automatically into horizontal rectangular bases of primitive roof models. In the second step, we determine the primitive roof model 
by its rectangular bases and the corresponding roof points. A primitive roof or a combination of roof primitives can then be 
assembled to a polyhedral roof model in the third step. The building can be completely reconstructed by projecting its roof boundary 
outline to ground. Because of the rectangulation process, our method has limitation in only describing the buildings with right angle 
corners. Nevertheless, under this assumption, the result shows that our methodology can successfully reconstruct buildings with 
complex roofs over Purdue University campus. We demonstrate the results with pertinent discussions. 
1. INTRODUCTION 
3-D building reconstruction has been an active research subject 
in computer vision, computer graphics and photogrammetry 
since 1980's. An efficient building construction can benefit fast 
data collection, support effective photorealistic visualization, 
and facilitate 3-D geospatial query and database creation. 
A common strategy for automatic building reconstruction is to 
utilize building model base. Three most popular model types are 
parametric models, CAD (Computer Aided Design) models, and 
generic models. In the parametric model (Lang and Förstner, 
1996; Fischer et al, 1998), buildings are described by a number 
of parameters, whose values are to be determined in the 
reconstruction process. This model has limitations in describing 
complex buildings, because the number of parameters is fixed 
and a complex building may have many levels/floors and 
complex rooftops with irregular shape. The CAD based 
modeling approach classifies buildings into different primitive 
components. Buildings must be described in advance with a 
fixed geometry and topology. Because of -such pre-definition 
procedure, the exploration of unknown or complex buildings is 
constrained by the lack of variability in this model. Generic 
models allow for variation in building structures, which 
indicates the number of geometric parameters is free. There are 
three subclasses of generic models, including prismatic models 
(Wenidner, 1997), polyhedral models (Griin and Wang, 1998) 
and constructive solid geometry (CSG) models (Giilch et al, 
1999; Zlatanova et al, 1998; Norbert and Brenner, 1998). 
However, the requirement on pre-defined models still limits the 
types of realistic buildings to be reconstructed. 
This paper is focused on using unstructured distinct points for 
3-D building reconstruction. Similar studies have been reported 
by using roof points. Griin and Dan (1997) proposed a topolog 
builder TOBAGO for semi-automatic building reconstruction. 
The method is a model-based approach that requires the 
operator measuring all roof points, however, no specific 
sequence needs to be followed. Each “roof unit” is a complete 
553 
point cloud to be proceeded individually. First, a K-Parser 
classifies the roof into six generic roof models based on the 
number of ridge points. Then, a G-Parser exploits the geometric 
criteria within a particular roof class to enclose the 3-D points 
as a complete CAD building model. However, TOBAGO might 
fail if no corresponding roof unit is found in the pre-defined 
model database. Griin and Wang (1998) proposed another 
generic topology generator CC-Modeler for CyberCity (CC) 
modeling. The data in CC-Modeler are regarded as 3-D point 
clouds manually measured by the operator. During data 
acquisition, 3-D point clouds for each building need to be 
divided and labeled into two groups: boundary points and 
interior points. Boundary points for building roofs are digitized 
by following certain order and interior points can be measured 
without sequence. For reconstruction, the data is treated by a 
consistent labeling algorithm based on probabilistic relaxation. 
In this paper, our effort is devoted to structuring the unordered 
distinct points into 3-D buildings. The 3-D point cloud is 
manually measured from a pair of stereo images. The 
measurement of 3-D point clouds for all roof corners must be 
complete, including any hidden ones. For this objective, we 
focus our attention on reconstructing the buildings only with 
right angle corners and propose a model-based approach to 
regularly construct these unstructured points. In this study, CSG 
modeling rule is applied that a complex building roof can be 
decomposed into several primitive CAD models. Combinations 
of primitives are created by Boolean operations using a CSG 
tree. One primitive or a combination of primitives can be 
combined to form a polyhedral building. This proposed 
reconstruction method is formulated as a process of finding 2-D 
rectangles, forming 3-D polyhedral primitives, and assembling 
them to a building. Presented in this paper are successfully 
reconstructed Purdue University campus buildings and their 
comparison with aerial images. Properties of the proposed 
approach and its further improvement are also discussed based 
on our experience.