Proceedings 18 International Symposium CIPA 2001 
Potsdam (Germany), September 18 - 21, 2001 
AUTOMATED ARCHITECTURE RECONSTUCTION 
FROM CLOSE-RANGE PHOTOGRAMMETRY* 
Tomas Werner; Frederik Schaffalitzky and Andrew Zisserman 
Department of Engineering Science 
University of Oxford 
Oxford, 0X1 3PJ, UK 
{wemer,fsm,az}@robots.ox.ac.uk 
http://www.robots.ox.ac.uk/~vgg 
KEY WORDS: 3D Computer Vision, Multiple View Geometry, Plane Sweeping, Inter Image Homography 
ABSTRACT 
We describe a method of automated reconstruction of buildings from a set of uncalibrated photographs. The method proceeds in two 
steps (i) Recovering the camera corresponding to each photograph and a set of sparse scene features using uncalibrated structure 
from motion techniques developed in the Computer Vision community, (ii) A novel plane-sweep algorithm which progressively 
constructs a piecewise planar 3D model of the building. In both steps, the rich geometric constraints present in architectural scenes 
are utilized. It is also demonstrated that window indentations may be computed automatically. 
The methods are illustrated on an image triplet of a college court at Oxford, and on the CIPA reference image set of Zurich City 
Hall. 
1. INTRODUCTION 
There has been intensive research effort in the Photogrammetry and Computer Vision communities on reconstruction of architecture 
from photographs. For example, the following large scale projects dealt with various degrees of automated scene recovery, generally 
starting from cameras with known calibration and/or range imagery: Ascender [4], Facade [14], IMPACT [1,6, 10], and RESOLV 
[13]. 
In particular, the Facade project demonstrated the high quality of models that could be constructed manually from photographs using 
a paradigm based on first constructing a polyhedral approximation of the scene and then considered deviations from this 
approximation. The aim of the work here is an automated Facade modeller - the goal at this stage is first to recover a piecewise 
planar model that approximates the dominant planes in the scene and their delineation; and then to use these planes to organize the 
search for perturbations from the plane such as indentations (e.g. windows) and protrusions (e.g. window sills). 
We are interested here in architectural scenes which typically contain planes orientated in three dominant directions which are 
perpendicular to each other, for example the vertical sides of a building and the horizontal ground plane. It is assumed that the scene 
contains three such principal directions and that the images contain sufficient information to obtain the vanishing points of these 
directions. 
We describe a method which proceeds in two steps: first, the cameras are determined from the images. We assume that the cameras 
have square pixels and determine the camera matrices using a combination of multiple view matching, and vanishing points 
corresponding to the three principal scene directions. This is described in section 2. 
The second step is to build the piecewise planar model given the cameras. Here we use a "plane sweeping" approach to determine 
the planes in the principal directions. This plane sweeping strategy builds on previous work on automated reconstruction from aerial 
views by Baillard et al. [1]. It is a powerful building block for an architecture reconstruction system enabling the main walls to be 
recovered efficiently and reliably. Once the main structures are determined, more demanding searches for smaller piecewise planar 
parts, details in the wall, etc. follow. This is described in section 3. 
M 
ÌSm* 
ì 
(a) 
«Wr • 
(b) 
(c) 
Figure 1: Three images of Merton College, Oxford, acquired with a hand held low cost Olympus digital camera. 
The images are 1024 x 768 pixels. 
This work was supported by EC Project VIBES and an EC Marie Curie fellowship.