INEXPENSIVE AND ROBUST 3D MODEL ACQUISITION SYSTEM FOR 
THREE-DIMENSIONAL MODELING OF SMALL ARTIFACTS 
Ula§ Yilmaz, Oguz Ôzün, Burçak Otlu, Adem Mulayim, Volkan Atalay 
{ulas, oguz, burcak, adem, volkan}@ceng.metu.edu.tr 
Department of Computer Engineering, 
Middle East Technical University, 
Ankara, Turkey, TR-06531 
KEY WORDS: acquisition, calibration, digitization, modeling, registration, triangulation, high resolution, three- 
dimensional, distortion, image, Internet, object, sequence, texture, web-based, photo-realism, virtual reality. 
ABSTRACT 
An image based model reconstruction system is described. Real images of a rigid object acquired under a simple but 
controlled environment are used to recover the three dimensional geometry and the surface appearance. Proposed system 
enables robustly modeling of small artifacts with high resolution geometry and surface appearance. Such artifacts may 
have handles, holes, concavities, cracks, etc. The proposed system enables robustly modeling of these properties also. The 
models obtained by this method are stored in Virtual Reality Modeling Language format which enables the transmission 
and publishing of the models easier over Internet. 
1 INTRODUCTION 
Realistic looking three dimensional (3D) models are im 
portant components of virtual reality environments. They 
are used in many multimedia applications such as 3D vir 
tual simulators, 3D educational software, 3D object databases 
and encyclopedias, 3D shopping over Internet and 3D tele 
conferencing. The virtual environments enable the user in 
teract with the models. The user can rotate, scale, even 
deform the models; observe the models under different 
lighting conditions; change the appearance (color, mate 
rial, etc.) of the models; observe the interaction of a model 
with the other models in the environment. 
The most common way of creating 3D models for virtual 
reality environments is manual design. This approach is 
very suitable for the creation of the models of non-existing 
objects. However, it is cost expensive and time consum 
ing. Furthermore, the accuracy of the designed model for 
a real object may not be satisfying. In fact, the number 
of vertices of a simple model varies between 10000 and 
100000, which is relatively high for manual design. In a 
second approach, geometry of real objects is acquired us 
ing a system that captures directly 3D data. Such systems 
are constructed using expensive equipments such as laser 
range scanners, structured light; touch based 3D scanners, 
or 3D digitizers. In most of these active scanning systems, 
the texture of the model is not captured while the geome 
try of the object is acquired precisely as a set of points in 
the 3D space. This set can then be converted to polygonal 
model representations for rendering (Soucy et al., 1996). 
In a third approach, the model of a real object is recon 
structed from its two dimensional (2D) images. This tech 
nique is known as image-based modeling. Even using an 
off-the-shelf camera, considerably realistic looking models 
with both geometry and texture is reconstructed (Niem and 
Wingbermuhle, 1997, Matsumoto et al., 1997, Schmitt and 
Yemez, 1999, Pollefeys et al., 2000, Mulayim and Atalay, 
2001, Yilmaz et al., 2003, Mulayim et al., 2003). 
Recent advances in computer vision in addition to pho- 
togrammetry make it possible to acquire high resolution 
3D models of scenes and objects. One of the most pop 
ular applications has emerged as digitizing historical and 
cultural heritance such as Digital Michelangelo and ACO- 
HIR (Levoy et al., 2000, ESPRIT, 1998). However, recon 
struction of a complex rigid object from its 2D images is 
still a challenging computer vision problem under general 
imaging conditions. Without a priori information about 
the imaging*environment (camera geometry, lighting con 
ditions, object and background surface properties, etc.), it 
becomes very difficult to infer the 3D structure of the cap 
tured object. For practical purposes, the problem can be 
simplified by using controlled imaging environments. In 
such an environment, camera makes a controlled motion 
around the object, and background surface and lighting are 
selected to reduce the specularities on the acquired im 
ages (Niem and Wingbermiihle, 1997, Matsumoto et al., 
1997, Schmitt and Yemez, 1999, Miilayim et al., 1999, 
Lensch et al., 2000, Ramanathan et al., 2000, Yilmaz et 
al., 2003, Miilayim et al., 2003). 
Figure 1: Image acquisition system consists of a turn table, 
a camera and a computer.