International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
(a) 
  
35 40 
(b) 
Figure 4: Visualization of orientations in the rectified frame: (a) image region with vectors visualizing the edge orientation 
(vector length corresponds to the magnitude). (b) histogram of edge orientations 
correct matches are the inliers resulting from computing 
the essential matrix. The experiment is carried out with 
two different versions for the support region: Version one 
uses the sector as described in our approach. In version two 
the support region is centered skew symmetric around the 
point of interest. This comparison assesses the increase in 
discrimination ability when using only one sector of the in- 
terests points surrounding. The inlier rate for our approach 
is represented by a solid line, the dashed line is the inlier 
rate for the skew symmetric support region. 
Figure 5 shows the differences in the used support region. 
Figure 6(a) shows the results for the turntable sequence 
     
  
dé P; 
15 E uy gs d ly 
Figure 5: Illustration of the two cases for the support re- 
gion. Left: support region lies inside the sector defined 
by the intersecting lines. Right: support region lies skew- 
symmetrically around intersection point 
for real images. One can clearly see the superior behavior 
of the sector representation (approx. 20 percent increase 
in performance). The variance can be explained by oc- 
clusion effects e.g. when a new face of the box appears 
and the number of possible candidates increases or when a 
face disappears and the number of candidates drops. Our 
approach outperforms the version with the skew symmet- 
ric support region is the rotation between the cameras in- 
creases. In Figure 6(b) illustrates the results for the syn- 
thetic turntable sequence. The scene consists of a planar 
object with several differently structured textures 'glued' 
on it. Due to the lack of depth discontinuities the perfor- 
mance between the two versions for the support region dif- 
1122 
fers less, which again nicely demonstrates the superiority 
of Zwickels on depth discontinuities. 
(a) (b) 
Figure 6: Illustration of the invariance against viewpoint 
changes: The rotation between the two cameras is in- 
creased in five degree steps form five to ninety degrees. 
The continuous line is the result for our approach, the 
dashed lines is for the centered support region. (a) shows 
the results for the data of the turntable sequence for real 
images. The variance results from occlusion effects, when 
new faces appear or other vanish. (b) illustrates the results 
for the synthetic turntable sequence. 
In the following experiment we took several image pairs 
and evaluated the matching performance. Figure 7 shows 
the 30 percent best matching correspondences for those 
image pairs. Table | lists the results for four different im- 
age pairs. Results using other images are similar. In col- 
umn 2 we list the number of total matches found, column 3 
shows the number of best matching correspondences used 
for estimating the epipolar geometry. In column 3 and 4 
we list the number of inliers and outliers accepted or re- 
jected by enforcing epipolar consistency. Note that all im- 
age pairs show a significant rotation between views. It is 
clearly seen that our novel method produces many good 
matches and only few outliers. The matching, including 
the estimation of epipolar geometry, takes between 6 and 
   
   
   
  
   
   
  
  
  
   
  
  
  
  
  
  
  
  
  
    
   
    
  
   
  
  
   
  
   
  
   
    
   
  
   
  
  
    
  
  
   
  
  
   
   
    
Interi 
  
Tabl 
as w 
mate 
seve 
regi 
14s 
We 
ant « 
thes 
as v 
fori 
Zwl 
elsp 
peti 
use: 
pari: 
ofe 
reco 
ACI 
This 
Gra: 
part 
grar 
the 
(AC 
to tl 
Kon 
quei 
REI 
Bau 
wid 
Can 
tion 
Inte 
Har 
edg