CIPA 2003 XIX th International Symposium, 30 September-04 October, 2003, Antalya, Turkey 
Definition 1 Let the extracted color values for a given par 
ticle be C = {co, Ci, ..cm~\] for an image sequence S = 
../jV—l }• The color of the particle for this sequence 
is photoconsistent if 
1. There exists at least two images in S in which the par 
ticle is not occluded. 
2. The particle is not on the background in any of the 
images in S. 
3. C a < © where C a is the standard deviation of inten 
sity values of the colors in C. 
Overall appearance reconstruction process is described in 
Algorithm 2: by projecting the particle on the source im 
ages, a set of candidate colors is extracted. If the recon 
structed set is photoconsistent then the color of the particle 
is selected as the median of this set. If a particle is occluded 
in all of the images or a photoconsistent color cannot be 
extracted from the sequence, there occurs regions whose 
appearance cannot be recovered on the model. The colors 
of the particles in these regions are interpolated using the 
colors of the adjacent particles. 
Algorithm 2 Recovering the color of a particle, 
reset the candidate color set of the particle to empty set 
for all images in the sequence do 
if particle is visible in the image then 
project particle on the image 
insert the extracted color in C 
end if 
end for 
if C is photoconsistent then 
set the color of the particle to the median of C 
else 
set the color of the particle to the color of the nearest 
particle whose color is consistent. 
end if 
4 EXPERIMENTAL RESULTS 
The experiments are performed on a personal computer 
with 512 MB of RAM, Intel PHI 800Hz CPU and 32MB 
frame buffer. The images are captured with a 2/3” Color 
Progressive scan CCD camera at a resolution of 1294x1030. 
While rendering the final model, texture mapping is per 
formed using the routines provided by OpenGL. 
As shown in Figure 6, the algorithms are successful in re 
moving the highlights which possibly occur at the image 
acquisition step. Reconstruction results given in Figure 7 
show that objects which have holes and arbitrary shapes 
can also be modeled by this method. Concavities can suc- 
cesfully be detected and carved in our method as shown in 
Figure 8 and Figure 9. 
Figure 6: (a) Sample images, (b) sample reconstructions 
for the “box” object. 
images 
Figure 7: Sample reconstructions for a “cup”object and a 
small artifact, “toy” object. 
models 
289