Shih-Hong Chio 
  
illustrates these two situations. 
Suppose that several points are on the 2-D line, a structure point is preferable than a junction point. Of course, the 
automatically locked corners might not be accepted by the operator. In that case, it can be unlocked by pressing one key. 
When the roof patch has been formed in the left image, 
the Point Database of the right image is then used to find 
the stereo correspondence of the corners defined in the 
left image. As mentioned in the previous section, when 
the system can’t obtain the 3-D coordinates of the 
corners from the information supplied by interactive 
selection and confirmation of lines and corners in the left 
image alone, it must search for the corresponding corner 
points in the right image in order to be able to use the 
forward intersection to calculate the object coordinates. 
Since we have confirmed the roof patch in the left image 
  
Roof Patch in the Left Image Roof Patch in the Right Image and the stereo images are normalized images [Cho and 
Shenk, 1992], it is easily to determine the search range in 
Fig.6. Diagram of Searching for the Corresponding Corner the right image based on the epipolar geometry. Possible 
correspondences can be collected along the epipolar line. 
Under our model assumption in Fig.2, at least one of the two adjacent edges of a corner must be horizontal in the object 
space. Based on the semantic information of the initial 3-D linear segment at least which one of these two edges must 
be a horizontal one can be determined. The direction of horizontal line from this corner to the neighboring corner is then 
determined. Since the direction of the corresponding horizontal line in the right image must be parallel to its stereo mate 
in the left image, its direction in the right image is also known. But the location of that line is of course still not known. 
If several candidate points near the endpoint of this horizontal line are available for a corner (cf. Fig.6), we have to 
determine which one shall be chosen. The decision is made on the condition whether a line parallel to that horizontal 
can be drawn from that point. Thus, from each candidate we can draw a line parallel to that direction. For each pixel 
along the line the gradient change across this line can be calculated. The line with the maximum gradient change has the 
greatest probability to be an edge. Therefore the point related to this line will be chosen as the corresponding corner. 
Point Database could also employed in the third case of interactive modification of the roof patch corners after a roof 
patch is reconstructed but the operator is not satisfied with the result. The operator drags the cursor near the corner 
which should be modified. By pressing a special key, the system will try to modify this corner by changing the 
corresponding point, similar to the procedure of finding the corresponding corner (cf. Fig.4). The operator examines 
interactively each result and decides whether the change is acceptable or not. 
4 EXPERIMENTS AND RESULTS 
Fig. 7 and Fig.8 illustrates the test stereo images and the results of the tests before and after the final interactive 
modification. The images cover an urban area very typical in Taiwan. They are scanned with 25 jim resolution. The 
original photos were taken by the normal angle camera with 60% overlap and their image scale is about 1: 6,000. Table 
1 lists the difference between the result from our system and the reference data measured carefully by an experienced 
operator on the Leica-Helava SOCET SET workstation. The number of complete roof patches without occlusion in 
these test stereo images is 21. 13 of them have been reconstructed by our system. Because of the failure of linear feature 
extraction due to weak contrast, five roof patches were unable to be reconstructed by our system. We compare our 
results to the reference on the location, the orientation and the dimension of the roof patches. From Table 1 we can see 
that no matter whether the reconstructed results are interactively modified by the operator or not in the final stage, the 
difference of plane position are less than 0.5m(3.3 pixels). Additionally, the difference of height is almost also below 
0.5m except at one location which is 0.60m(4 pixels). Although it shows the 3.3 pixels in position difference and 4 
pixels in height difference respectively, however, the differences could be negligible for the 3-D applications. As for the 
difference of orientation of roof patch, most of them are less than 6 degrees. Only three patches are larger. The last two 
columns show the difference of areas. No matter if the area could be a judging factor, we can see that the average 
percentage of the difference in areas is below 10 %. 
We believe that the accuracy of the result is more than enough for some applications like city planning, 
telecommunication and so on. Although it is hard to see the fine differences after interactive modification in the final 
stage from this test, we still believe that the interactive approach will give better results to some extend. 
5 CONCLUSIONS 
This paper presented an interactive system for roof patch reconstruction based on 3-D linear segments by integration of 
the interpretation ability of the operator. The reconstruction results are very promising. Our paper also introduces and 
  
188 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.