AUTOMATIC EXTRACTION OF BUILDING ROOFS FROM PICTOMETRY’S 
    
ORTHOGONAL AND OBLIQUE IMAGES 
Yandong Wang 
Pictometry International Corp. 
Suite A, 100 Town Centre Dr., Rochester, NY14623, the United States 
vandong wang@pictometrv.com 
Commission III, WG III/4 
KEY WORDS: Photogrammetry, Extraction, Building, Image, Edge, Automation. 
ABSTRACT: 
Automatic extraction of buildings from digital images aims at detection of buildings from digital images and reconstruction of roof 
structure automatically. At Pictometry, more than 30 million images are captured every year and how to extract the useful 
information of objects on the ground from the existing image library for various applications is a big challenge we face now. In this 
paper, an automatic approach for extraction of building roof from digital orthogonal and oblique imagery is proposed. The proposed 
method uses image processing technique to derive the accurate 3D structure of building roof for accurate roof measurement, 3D 
modeling, computation of building footprints, etc. It consists of three major steps, i.e. extraction of roof corner and ridge points from 
the images, automatic matching of roof corner and ridge points between orthogonal and oblique images and grouping of the matched 
roof points to create roof facets. In this study, the modified Moravec operator is used to extract feature points from digital images. To 
find roof points which cannot be extracted by the point extractor, edge information is also extracted. Due to the nature of roof points, 
especially corner points and the difference between orthogonal and oblique images, a feature-based image matching technique is 
used to derive 3D information of roof corner and ridge points, instead of area-based matching. To match roof points correctly, edges 
associated with a corner or ridge point and their properties are used. After 3D roof points are generated, roof points belonging to the 
same roof facet are grouped together by using their spatial relations. Once points belonging to the same facet are found, a surface is 
fitted to the points and outliers can be removed during this process. 
1. INTRODUCTION 
3D modeling of buildings has many applications in different 
areas such as 3D city modeling, communication, insurance, 
urban planning, etc. Automatic extraction of buildings from 
digital images has received significant attentions from both 
computer vision and photogrammetry over decades and there 
are a number of methods developed by researchers in these 
fields. The early research was focused on the extraction of 
buildings with simple structures (rectangular shape) (Huertas 
and Nevatia, 1988; Fua and Hanson, 1991; Dang et al, 1994; 
Roux and McKeown, 1994; Lin et al, 1995). Some approaches 
dealing with buildings with complex structures were developed 
in Lang and Fórstner (1996), Henricsson (1996) and Taillandier 
and Deriche (2004). Recently an approach for extraction of 
building facade has been developed by Xiao et al (2010). 
Pictometry started to capture geo-referenced imagery using its 
proprietary imaging system more than a decade ago. Currently 
more than 30 million of both vertical and oblique images are 
captured every year at Pictometry and the number still increases 
every year. One big challenge we face is how to extract useful 
information from the existing image library for different users. 
In this paper, an approach for automatic extraction of buildings 
from both vertical and oblique imagery is presented. The 
approach focuses on the reconstruction of building roof using 
image processing methods for different applications such as 3D 
city modeling, insurance and urban planning. It consists of 
three major steps, i.e. automatic extraction of features and 
generation of topological relations between features, matching 
of 2D features to derive 3D features and grouping of 3D 
features to generate roof facets. In the following sections, a 
building model for automatic building extraction will be 
described in section 2. The details of feature extraction, 
matching and grouping will be given in section 3. Some testes 
results will be given and discussed in section 4. 
2. BUILDING MODEL AND EXTRACTION 
STRATEGY 
Various building models have been developed for building 
extraction based on the type of buildings to be extracted and the 
resolution of images to be used. They range from using simple 
geometric constraints such as rectangular shape to the use of 
complex 3D geometric constraints for extraction of buildings 
with complex structures. Extraction strategy largely relies on 
the model to be used for building extraction. It can be data- 
driven, model-driven or hybrid. In a data-driven strategy, 
extraction usually starts from extraction of generic features 
from imagery such as points and lines. 3D features are 
generated by matching 2D features from overlapping images 
and buildings are reconstructed by grouping 3D features. À 
review on building models, building extraction strategy and 
performance of some existing methods can be found in Mayer 
(1999). 
The model used in this study uses 3D geometric constraints. It 
is assumed that a residential building roof consists of a number 
of facets which connect to each other. Each roof facet has a 
number of roof points which are connected by roof lines such 
as eaves, ridge and valley lines. A bottom-up strategy will be 
used for reconstruction of building roofs in this study. It starts 
with the extraction of point and line features from 2D imagery 
  
prir 
con