Proceedings 18 th International Symposium CIPA 2001 
Potsdam (Germany), September 18 - 21, 2001 
A TOP-DOWN APPROACH TO INTERACTIVE 3-D MODELLING 
OF ARCHITECTURAL SITES 
Philippe Even 
Atomic Energy Commission, Robotics and Interactive Systems Laboratory, (CEA/SRSI) 
PO Box 6, F-92265 Fontenay-Aux-Roses, France 
Since September 2001, the author is at LORIA, BP 239, 54506 Vandœuvre lès Nancy, France, e-mail:Philippe.Even@loria.fr 
KEY WORDS: 3-D Modelling, Architectural Photogrammetry, CAD, User Interfaces, On-line 
ABSTRACT 
Most 3-D model reconstruction techniques using video images start from 2-D features extraction, and then progress towards 3-D 
interpretation until the 3-D structure is fully recovered. In order to fulfil its specific needs for telerobotics applications, CEA designed 
a new paradigm based on a top-down process. The user first builds a mental representation of what he sees in the images. 3-D 
primitives are then superimposed on the image and interactively fit to the visual features. The basic idea is to reduce the number of 
manual tasks and also to enhance interactivity through immediate and comprehensible feedback in reply to the operator's actions. 
This principle was developed and evaluated through the Pyramide software. Although many functions are dedicated to robotics 
purpose (exploitation of robot joint values to estimate the camera position, low attention paid to the model visual realism), it includes 
also worth interesting features for other applications. Potential advantages for architectural sites modelling have been evaluated using 
the Zurich City Hall reference data set. This top-down approach was quite successful to deliver a structured and detailed model of the 
test building using only four views. 
1. INTRODUCTION 
Despite of considerable improvements achieved these last years in the computer vision community, full automatic 3-D modelling of 
large sites from images still requires important time and generally a lot of images to facilitate features matching. It depends on 
parameters that may be difficult to estimate, such as image quality or richness of the modelled objects texture, and inherently leads to 
a huge surface description of the scene, not always well adapted to the task to be performed. 
Therefore present solutions are mostly based on a manual approach. The level of detail is at the operator's appreciation so that the 
obtained model features relevant information with respect to assigned goals. The operator decides of what has to be modelled and 
with which precision, choosing for instance to place a surrounding box around some obstacle or to acquire a finer description of each 
object. When goals are clearly assigned, the model structuration and semantical contents are also suitably adapted to the task. Based 
on the human perception system capabilities, a manual mode ensures robustness to bad viewing conditions. Artefacts such as 
occluding objects are naturally ignored. 
In order to facilitate and thus speed up the modelling process, manual systems tend to integrate more and more interactivity. This 
interactivity is characterized by a close cooperation between the system and the operator through efficient assisted modes and also by 
immediate and comprehensible feedbacks in reply to the operator's actions. 
This paper highlights the advantages of a top-down approach to the interactive modelling of architectural environments, in which a 3- 
D model is directly realized and superimposed on 2-D images in order to ensure the consistency with real data. This approach is 
implemented into the Pyramide interactive modelling software that has been developed at CEA for telerobotics applications in hostile 
environments (Even and Mareé 1988). Particular attention is paid to the views exterior orientation problem, which constitutes the 
major difference between both application fields. This tool was tested in architectural context using the Zurich City Hall reference 
data set (Streilein, Grussenmeyer and Hanke 1999). A structured and detailed model of the building was delivered using only four 
views. Achieved results highlight this approach particular efficiency to exploit the high degree of structuration available in most of 
architectural sites. 
The top-down approach is exposed in next section, emphasis being put on its contribution to interactivity enhancement. 
Implementation within Pyramide system is detailed in section 3. The advantages for robotics purposes and the difference from 
architectural photogrammetry applications are outlined in section 4. Then section 5 focuses on interactive solutions to the objects 
orientation estimation problem. Methods based on visual features interpretation techniques are exposed. Section 6 relates the 
experiments held using the Zurich City Hall data set, with recorded time measured during the modelling work and views of the 
registered model over the exploited images. A conclusion and promising continuations to this work are given in section 7. 
* After September 2001, the author will be at LORIA, BP 239, 54506 Vandœuvre lès Nancy, France, e-mail:Philippe.Even@loria.fr