ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS", Bangkok, May 23-25, 2001 
252 
Fig. 11 The Result of Line Segmentation (Haar’s Wavelet) 
Fig. 12 The Result of Edge Matching (SUSAN Operator) 
Fig. 13 The Result of Edge Matching (Haar’s Wavelet) 
6.3 Stereo Matching with CEM 
To verify the influence of balance between competition and 
consensus process, comparison experiments was conducted. 
Fig. 14 shows the mapped image and shift vector from the left 
image to the right one by CEM with standard processing 
parameters. The depth map is also illustrated, which is 
calculated by shift vector and overlaid onto the left image. Fig. 15 
shows the result of a case when consensus operation is 
dominant compared to competition operation. It is noticed that 
enhanced consensus operation produces smoother shift vectors, 
yet overly strong application is not suitable for urban area. 
Fig.16 shows the images processed by Cl operation with Eq.2 
and Eq.3. The results processed by CEM with these images are 
illustrated in Fig. 17. Fig. 16 shows that Cl operator enhances 
areas where brightness change is strong, especially for the 
edges of buildings. With regards of edge parts, Cl operator has 
similar effect to the sum of absolute value of first-degree 
differential calculus in local area. By comparing Fig. 17 with 
Fig.14, it is observed that the mapping result with Cl operator 
gives almost acceptable depth map but leads more deformation. 
This result is caused by the areas of flat textures formed by Cl 
operator, which consequently causes mismatching by local 
minimum. Therefore when using Cl operator, techniques such 
as using only characteristic region for matching and interpolating 
other areas should be considered. 
Fig. 18 is the result of CEM with matched edge as constraints. 
Conjugate edges detected by SUSAN operator are used. In 
comparison with result in Fig. 14, it is observed that shifts of 
parallax are improved at neighboring area of buildings. However 
at some of area, local disturbances have occurred because of 
consensus operation, which tries to decrease abrupt parallax 
shift. This is because that CEM process with edge constraint 
cannot deal only with edge segments. It also has some effects 
on other areas. In the next step of this study, perfect separation 
of process between edges and other areas must be 
investigated. For example, at the first stage, match only edge 
segments with CEM. In later stages, match adaptively the areas 
between edge segments. 
(a) Mapped Left Image (b) Original Right Image 
(c) Shift Vector 
(d) Depth Map Projected to Left Image 
Fig. 14 The CEM Result with Average Parameters