hed . In other 
  
ching method. 
's of buildings. 
or optical axis. 
ce of building) 
[(x,y) on this 
| shape formed 
etely, template 
atched value : 
have adequate 
y) (1) 
nt in all feature 
iin in searching 
l < X delta T 
Je, 1991 )(Jianbo 
nage sequence: 
solution image 
yrmation model 
ted as follows. 
high-resolution 
L) 
2) 
V) 
Isao Miyagawa 
The high-resolution images can be captured about 2 [fps], the video images can be captured 30 [fps]. The trajectory from 
feature points on the high-resolution images is discrete. This trajectory can be completed by trajectory from feature points 
in the video images. 
  
  
  
Time-redundancy 
from Video Image 
  
  
  
  
N 
  
  
  
  
  
  
  
  
  
  
Feature Points 
Hybrid Feature Points 
composed from High-resolution images 
composed from High-resolution images 
and Video Images 
Figure 4: Trajectory of Feature Points 
Moreover, this transformation is effective for texture mapping. When the operator set feature points on a initial high- 
resolution image, texture data corresponding to the top surface of acquired 3D object shape is cutting automatically from 
video image, and fitting to acquired 3D object model. 
4 SHAPE RECOVERY WITH FACTORIZATION METHOD 
We have already developed an improved factorization method that utilize sensor information (Miyagawa, 1999)(Miya- 
gawa, 2000). Sensor information, i.e. yaw, pitch, and roll rotation can be used to create the camera motion matrix [M]. A 
registered measurement matrix [A]* composed from 2D feature points is decomposed using this camera matrix into shape 
matrix [5]. Here, [A]* & [U1][W3][Vi] using the rank-3 property of singular value matrix. In this paper, we expanded this 
method to fit hybrid feature points. 
[5] 2 «qi [p a" (MD) [Wa] [A] ed 
Moreover, 3D object shape can be placed within the unique coordinate system that we call Tokyo Datum. 
4.1 3D RECONSTRUCTION ON DELTA POINTS 
It is important to reconstruct 3D object models on the delta planes, because the height of each object is determined as 
the distance between points on the building’s roof and the delta plane. However, it is difficult to recognize the height of 
acquired 3D delta points using the factorization method, without the orthometric height data on the ground. It is shown 
in Figure 5(left side) that each 3D object model is reconstructed on a flat plane. This flat plane has no relationship to sea 
level. If delta points can be given 3D coordinate value from the orhtometric height data, we can make good use of the 
delta plane as a slope plane. In this case, each 3D object model can be reconstructed relative to sea level (Figure 5(right 
side)). We can reconstruct 3D object models on any delta plane if the delta points have orthomeoric height data on the 
ground. 
Î 4 
Z Value 3D Object 
3D Object 
Delta Plane Delta Plane 
Delta Point Delta Point 
  
  
^ Sea Level + Sea Level 
  
Figure 5: 3D Reconstruction on Delta Plane 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 611