The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part Bib. Beijing 2008 
684 
Figure 6. Projection of the LiDAR points for height Fitting. 
4. EXPERIMENTS 
A small urban area of Taipei City about 500hectare is selected 
for testing. The 1/1000 scale digital topographic maps have 
been pre-proceeded to generate building polygons. The grid in 
terval of the corresponding DEM is 4m. The point density of 
LiDAR point cloud is about 10 points within 1 m 2 , which is 
good enough to reconstruct normal building roofs. The aerial 
photos are taken by the Vexcel UltraCam D photogrammetric 
camera. The focal length is 101.4mm, the image size is 
7500*1 \500pixel, and the size of a pixel is 9*9pm. The average 
flight height is about 1930m, so the ground resolution is about 
0.17mlpixel. Meanwhile, we develop a PC program by C++ 
language to implement the proposed building reconstruction 
procedures. The interface is illustrated by Fig. 7. The operating 
sequences are as follow: (1) observe the topographic map in the 
left window and select the appropriate model; (2) click vertices 
V/, v 2 , and yj in sequences on the topographic map to give initial 
parameters; (3) examine projections on photos and adjust the 
model parameters if needed; (4) click the Fitting button to im 
plement LSMDF of plane and height optimal fitting; (5) exam 
ine projections on photos and adjust the model parameters if 
needed; (6) output and save the model parameters. A model is 
usually reconstructed within a minute, but the time for a build 
ing depends on its complexity. 
Figure 7. The program interface of MBBR. 
Figure 8. A complex building model reconstructed by box and 
polyhedral prism models. 
Our system currently provides three kinds of model for recon 
structing most of the modem buildings: box, gable-roof, and 
polyhedral prism model. A building model is composed of sev 
eral primitive models. Fig. 8 shows an example of a complex 
building reconstructed by box and polyhedral prism models. 
For the whole test area, two operators worked for one week and 
totally reconstructed 4130 buildings. Fig. 9 shows a part of the 
reconstructed city model. We select 30 buildings for correctness 
and accuracy evaluation. These models are first evaluated in 
their shape with aerial and terrestrial photos by human eyes. 
The correctness rate is about 88.5%. Then, the vertices coordi 
nates of the 30 building models are calculated from model pa 
rameters and then compared to the photogrammetric and ground 
survey result. Table 1 lists the statistics of the coordinate differ 
ences. The larger X-Y differences most due to the mismatch 
point, while the larger Z differences most due to the parapets. 
Table 1. Statistics of coordinates differences. 
Coordinates Differences 
AX 
AY 
AZ 
Average(m) 
0.051 
0.110 
-0.0146 
Avg. of Absolute Values(m) 
0.236 
0.294 
0.8816 
Std. Deviation (m) 
0.2953 
0.3490 
1.1400 
Figure 9. A part of the reconstructed 3D building models. 
Resulting from our experiments, most of the modem buildings 
can be modeled smoothly, and fitting result achieves the photo 
grammetric accuracy. However, some of the buildings are ille 
gally reconstructed into an arbitrarily shape, which makes it dif 
ficult to be modeled by our pre-defined model. In that case, the 
building should be decomposed into several parts for fitting and 
then aggregate into one composite model. For some traditional 
Chinese architecture, it is very difficult to reconstruct their 
curvy eaves by our pre-defined model. Fig. 10 shows an exam 
ple, in which the building can only be reconstructed approxi 
mately. 
Figure 10. Curvy eaves can only be reconstructed approxi 
mately. 
5. CONCLUSIONS 
The floating model is proposed as a model-based building re 
construction approach, which is a flexible 3D measuring tool