2-1 Distance between initial location/shape of the model and edge location in the image
The building whose shape and position in the existing map agree well to edges in the image have higher rate of success of the fitting. Figure 2-1 shows typical examples. On the other hands, the fitting fails for a rebuilt building as shown in Figure 2-2, which suggests that newly built buildings after a building is demolished can be detected. Here, it can be concluded that agreement of the shape and the position of existing building boundary data and the image is important for the model fitting, though further experiments are needed for the quantitative evaluation. Since the IKONOS image used in this research is an oblique image, the height of the building has caused horizontal shift of the building boundary data which results in the failure of the model fitting.

In the following figures, the existing residential map data is drawn by blue line, the result of processing with SNAKE is drawn by blue line connected in a blue point.

- Figure 2-1: initial value and image: success example





-Figure 2-2: initial value and image: failure example-





2-2 Contrast of edges
If buildings have edges with good contrast and connectivity in the edge image, model fitting tend to succeed. On the other hands, it is likely that the model fitting fail for the buildings where the edge contrast is low and the connectivity of the edge is poor. Especially, when there are trees adjacent to the building or shadow on north side of building, they are recognized as parts of the building, and SNAKE edges tend to be caught by those other edges. However, there is a tendency that a good result is obtained easily when roof color is white. It is due to stronger contrast of white roofs to building shadow and ground color. In addition, the experiment suggests that distance to the adjacent buildings also influences. Figure 2-3 and Figure 2-4 show the edge map in the building in Figure 2-1 and Figure 2-2. When the edge contrast and connectivity is good, SNAKE edge can follow neatly along the edge.

- Figure 2-3: Edge image: Good contrast-


- Figure 2-4: edge image: Poor contrast and connectivity-


2-3 Three dimensional complexity
In case of deck, model fitting usually succeed, because the roof has homogenous color, while gable roofs may not be successfully extracted, because north side of the roof are regarded as the shadow of the buildings.

In case of a building consisting of different structures (composite building), the height of the building roof is not homogenous, which generate edges with complicated shapes not consistent with the building boundary data in 2D maps. Typical success and the failure example are shown in Figure 2-5 and Figure 2-6.

- Figure 2-5: failure example-


-Figure 2-6: success example-


2-4 Iteration number
It was examined whether there is a correlation between success/failure of the fitting and iteration number. Iteration number is total times that points move with SNAKE.

When targeted building is quadrangle (node number is 4), there is a tendency that the fitting succeeds if the iteration number is ten or less, but there is a tendency that the fitting fails when it was a lot of from ten. This value might be helpful in the classification of the success or the failure o the fitting.

3 Dependency on initial values
Especially, for the building where the fitting does not succeed in Chapter 2, the initial value in terms of the distance with "Correct answer" is changed to see how or whether the fitting results fail. Thus the initial value dependency of the fitting result is analyzed. Here, the initial value of the vector data is recreated along the rooftop side on the building in the IKONOS image.

Moreover, the similar procedure is applied to the buildings not recorded in the existing residential maps, that is, new buildings. As for SNAKE, the initial value dependency is thought to be high as a result. The following two examples are typical examples of showing the dependency.

3-1 For edge map
If the initial value is not appropriate, the fitting does not often succeed. When the building is overcrowded or the trees are grown in surroundings, SNAKE is pulled to the edge of surrounding features. However, when the initial value is given on the outline in the building again, a good result is obtained. Figure 3-1 shows an example of giving the initial value, and the success in the fitting.

- Figure 3-1: building overcrowdedness and success example-

-The first fitting results-	-Second fitting trial by changing the initial values-

3-2 For high building
An oblique photograph caused the shift in the location of the building roof shape. The fitting is able to have done well when there is a difference in the color of the wall and the rooftop.

Oppositely, when the color of the wall and the rooftop are almost the same, the fitting usually fails. It is thought that this is due to the poor contrast. So, when the entire wall is shadowed, the fitting can succeed. Figure 3-2 shows horizontal displacement of the initial value, and the example of the success of the fitting changing the initial value.

-Figure 3-2: high building-

-The first fitting results-	-Second fitting trial by changing the initial values-