in 
he 
in- 
int 
Fa- 
ty 
;t. 
rs 
y- 
th 
LC- 
vn 
w- 
d, 
The left boundary of the building model is modeled correctly 
from laser points by intersecting two large wall planes. How 
ever, in practice it is matched to a strong contrast caused by a 
water pipe on the wall (see Figure 8(b)). This again, reveals the 
limitations of this refinement method. 
6.3 The wall with high windows 
(b) Before refinement 
In this example, the refinement is applied to improve the windows 
extracted from the holes from laser points of a wall (Pu and Vos- 
selman, 2009). The contrast of a window and its surrounding wall 
are usually rather obvious in optical data. Strong line features are 
frequently found at the windows’ boundaries and frames, and can 
be used to refine the information from the laser altimetry. Figure 
9(b) shows the window rectangles extracted from laser points, 
which contain a lot of errors due to limitation of segmentation and 
modeling algorithms. In Figure 9(a) we apply the same match 
ing stretchy for refinement purpose, and the final result in shown 
in Figure 9(c). Most windows' boundaries are well corrected ac 
cording to the image lines. The second left window in the upper 
row is not improved, because the difference between the modeled 
shape and the actual shape is too large to correlate them. There 
are some remaining errors, such as the first, third and sixth win 
dow (from left to right) in the lower row. This is because the 
parameters of Hough transform are too strict to generate any can 
didate line. 
The textured final model after manual adjustment is shown in Fig 
ure 10. Without the refinement, 42 vertices need to be manually 
adjusted. Only 10 vertices need to be adjusted after the refine 
ment. Note that the window rectangles are intruded inside the 
(c) After refinement 
Figure 9: Matching and refining window boundaries 
6.4 Summary 
The effectiveness as well as limitations of our refinement method 
are examined through the three test cases. We realize that the 
refining effect relies on the following prerequisites: 
• Accurate exterior and interior orientations. In particular, the 
selected tie points for spatial resection should be sufficient 
(four or more), and should be distributed equally in both hor 
izontal and vertical directions to minimize the computation 
error. 
• No large occlusions in front of the building facade. 
Figure 8: Matching model edges with image lines for refining the 
town hall’s model 
(a) Raw image 
.»-.V*»' « ** 
(b) Matching