International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
(c) (d) (e) 
Figure 14. The arbitrary-view images of man using texture 
mapping and 3D model rotating scheme. (a) Upper left view 
image, (b) Upper right view image, (c) Left view image, (d) 
Center view image, (e) Right view image. 
  
NO. n M 
Figure 10. Disparity map. (a) Block-Matching algorithm, (b) 
Population-Based Incremental Learning algorithm, (c) Proposed 
algorithm without object extraction, (d) Disparity map by using 
proposed algorithm. 
   
Figure 15. The arbitrary-view images of man using texture 
mapping and 3D model rotating scheme. (a) Upper left view 
image, (b)Upper right view image, (c) Left view image, (d) 
Figure 11. Disparity map of "claude" stereo image. (a) Block- e : e 
e pantvimap ZO Center view image, (e) Right view image. 
Matching algorithm with fixed-size window(7x5), (b) 
Population-Based Incremental Learning algorithm, (¢) Proposed 
algorithm without object extraction, (d) Disparity map by using 
proposed algorithm As the Figure 10(a) and Figure 11(a) indicate, since matching 
) g ; 
information doesn't exist in background, mismatching (noise) 
phenomenon is dominant. Also mismatching result occurs in 
projective distortion region and occluded region. The 
PBIL(Population Based Incremental Learning) (6) contains 
neighborhood considering characteristics, noise phenomenon is 
reduced. But mismatching phenomenon is much increased in 
the projective distortion region and occluded region(Figure 
10(b) and Figure 11(b)). Figure 10(d) and Figurel1(d) are the 
final result of proposed stereo matching algorithm. The result 
| shows that considerable improvements are obtained especially 
(a) (b) in projective distortion region. And other mismatching 
problems, i.e. noise, luminance, boundary problem, is reduced. 
Figure 12. Isometric plot of the disparity maps computed by ^ [n order to demonstrate performance of algorithm suggested, we 
proposed method. (a) Down left view, (b) Down right view. generate the texture mapping images (Figure 14 and 15). These 
different view images are generated by 3D model rotating 
scheme in the different viewing angle. Figure 12 and Figure 13 
show that proposed algorithm is superior to conventional 
algorithm, i.e. reduced noise, increase of matching reliability in 
the face (especially nose, cheek, and hair area). 
  
  
(a) (b) S. CONCLUSION 
A new stereo matching approach using probability-based 
window  warping algorithm was presented to improve 
conventional stereo matching method. Since the projective 
Figure 13. Isometric plot of the disparity maps computed by the 
proposed method. (a) Down left view, (b) Down right view. 
772 
AM a LL