CIP A 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
i, i + 1 and the voxel’s photoconsistency value is calcu 
lated using texture similarity measures among the projec 
tion regions on the images i — 1, i, i -f 1. Then, the voxel 
with maximum photoconsistency value is found and all the 
voxels existing between this voxel and the camera center 
Ci loose votes in an increasing order as they become closer 
to Ci. This process is repeated for all the rays which can be 
generated from the view i. When the process is performed 
for all the views, excess voxels caused by the silhouette 
based reconstruction lose most of their initial photocon 
sistency votes. Then by thresholding, this excess volume 
is carved. Figure 3 explains better the idea. In this figure, 
darkest colored voxels get the highest voting; i.e. they have 
the maximum texture similarity according to the algorithm. 
The color of voxel shows its vote. 
Figure 3: Darkest colored voxels get the highest voting. 
3 APPEARANCE RECONSTRUCTION 
There exist several studies for appearance reconstruction 
of 3D models from real images (Niem and Broszio, 1995, 
Gen§ and Atalay, 1999, Lensch et al., 2000, Neugebauer 
and Klein, 1999). In most of these studies, the model is 
represented as a triangular wireframe, and each triangle is 
associated with one of the images for texture extraction. 
The method causes discontinuities on the triangle bound 
aries as shown in Figure 4, since adjacent triangles can 
be associated with different source images. Applying low 
pass filter on the boundaries cannot come up with a global 
solution. 
In this study, 2D texture mapping is used but to reduce the 
drawbacks due to the lack of third dimension information, 
the concept of surface particles is adapted (Schmitt and 
Yemez, 1999, Szelisky and Tonnesen, 1992). An abstrac 
tion is done on the actual representation of the model: the 
model is considered to be a surface composed of particles 
with three attributes: position, normal and color. While 
reconstructing the appearance of the model, instead of as 
sociating triangles to images, particles are associated with 
images for texture extraction as shown in Figure 5. Each 
particle’s color is extracted from the images independently. 
This is what makes the proposed method superior to the 
Figure 4: (a) Discontinuities on the triangle boundaries; 
(b) reconstruction using particles. 
others: since a triangle is not necessarily textured from a 
single image, there are not discontinuities on the triangle 
boundaries due to the fact of being textured from different 
images. Each particle on the surface is associated with a 
pixel on the texture map, and the color information of the 
texture map is recovered by this means. 
A particle is not necessarily visible in all of the images in 
the sequence, it can be occluded or a back face in some 
of them. So, the extraction takes place at two steps: vis 
ibility check and color retrieval. Visibility check is per 
formed using the particle normal and the particle position 
by simple hidden surface removal and occlusion detection 
algorithms. The particle is projected on the source im 
ages in which it is visible, and a set of candidate color 
values, C = {c 0 ,..cm-i} are collected. In this study, 
the candidate color values are fused in order to produce 
the most photoconsistent appearance. Before assigning a 
color value to a particle, it is decided whether the informa 
tion extracted from the source images is photoconsistent 
or not. The photoconsistency is defined in Definition 1. 
The value of 0 is empirical. It is expected that the val 
ues of a photoconsistent set concentrates around the view- 
independent color of the particle. This method is very suit 
able for removing illumination artifacts as shown in Fig 
ure 4-b. However, if the geometry of the object is not con 
structed precisely, the photoconsistency criteria will fail for 
most of the particles, which will cause irregularities on the 
surface appearance.