1e 
of 
Shih-Hong Chio 
  
INTERACTIVE ROOF PATCH RECONSTRUCTION BASED ON 3-D LINEAR SEGMENTS 
Shih-Hong CHIO', Shue-Chia WANG' and Bernhard WROBEL" 
"Department of Surveying Engineering, 
National Cheng-Kung University, Taiwan, R.O.C. 
e-mail: chio@www.sv.ncku.edu.tw and scwang@www.sv.ncku.edu.tw 
“Institute of Photogrammetry and Cartography, 
Darmstadt University of Technology, Germany. 
e-mail: wrobel@gauss 
  
ABSTRACT 
An interactive system for roof patch reconstruction is introduced in this paper. The reconstruction is based on 3-D linear 
segments. The concept of interactive approach combines the interpretation ability of human operator and the 
computation ability of the computer together to increase the successfulness of the reconstruction. To increase the 
successfulness of the automatic finding of roof corners, a specially designed Point Database is used by the system. The 
interactive approach is accomplished by human intervention in various steps during the process. Experiments show that 
the system is very practical. 
1 INTRODUCTION 
The general procedure for roof reconstruction could be thoroughly divided into three subtasks: the geometrical feature 
extraction, the detection of possible locations of roofs or roof boundaries, and the final roof reconstruction. In all three 
subtasks intelligent interpretation is needed more or less. Despite the rapid development of the computer technology, it 
is still a great problem to make the computer deal with the interpretation task correctly. The central point lies on the 
lack of intact theory to handle this problem. Therefore methods for the fully automatic reconstruction of buildings found 
in the literatures [Collins et al., 1995;Fischer et al., 1997; Henriccson and Baltsavias, 1997; Lin et al., 1995; McGlone 
and Shuffel, 1994; Nevatia et al, 1997; Shufelt 1996] can't meet the general requirements encountered in the practice. 
Hence, the reconstruction problem becomes another issue as how to effectively integrate the interpretation ability of 
operator into the system in order to improve the performance. The development of the so-called semi-automatic system 
for roof reconstruction [Gruen,1998; Gülch, 1997; Heuel and Nevatia,1995; Hsieh, 1995; Lang and FÓrstner,1996] is a 
substantial improvement for the correctness of the building reconstruction. However, semi-automatic approach handles 
the problem in a simple forward way. Once the interactively provided information from the human operator is fully 
used, the system prompts a solution. If the solution is wrong, there is hardly any further intervention from the operator 
possible. 
Therefore, in this paper we would like to present an interactive approach for the roof patch reconstruction on the basis 
of 3-D linear segments. It is called “interactive” because in addition to including the high interpretation ability of the 
human operators, we have designed a constant interaction between the operator and the computer. Therefore the 
difference between interactive and semi-automatic systems is that our interactive system draws stronger human 
intervention in various steps during the reconstruction procedure and that the result of intervention is immediately 
prompted by the system on the screen for the operator to confirm or deny it. If it is not a correct one, the operator has to 
interfere immediately. All the interactive actions take place mono-scopically in single image, no precise stereo 
measurement is needed. 
We have specially designed this trial and error approach for handling the very dense buildings in most urban areas in 
Taiwan. It is very common in Taiwan that people build extra illegal superstructures on the top of the legally constructed 
buildings. These illegal superstructures are most of the time in the form of gable roofs with very different sizes and 
heights adjacent to each other (cf. Fig.7 or 8). Since almost no side walls of the buildings could be seen on the image, 
only roofs can be reconstructed. 
Although lines or edges are the most extracted information from aerial images for building reconstruction, they are most 
of the time incomplete and not reliable. Because they are most of the time merely an accumulation of individual edge 
pixels without any semantic information in the object space. Therefore in our system the edges are used only as 
intermediate elements for determining roof corners and a roof patch is actually reconstructed by corners. For better 
usage of corner points, we have designed the Point Database for managing and searching the corners. 
The system first tries to find as much as possible all “meaningful” 2-D and 3-D line segments from the stereo images 
automatically without any human intervention. By “meaningful” we mean that a line segment, which due to its semantic 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 183