(a) (b) 
  
  
  
  
  
(d) 
  
  
(c) 
  
ge 
E 
(e) 
Fig 2: Test Cube. (a) Left image, (b) Interlaced image, (c) Right image, (d) contours of Q over the search area for one 
of the test points together with the epipolar line, (e) Surface produced from Q factor for a test point. As expected, Q 
gets lower values near to match point. 
Even better, the above filtering technique and the arithmetic 
operation have already been implemented in terms of speed 
optimization in most CPU chipsets. For example, Intel claims 
that blurring an image and subtracting it from another one of 
equal size needs about 2.92 clocks/pixel/plane. 
Furthermore, in the classical approach, the correlation 
procedure is followed by least squares matching. The later is 
even more complex in terms of arithmetic operations and thus 
results to even longer execution time. 
4. EXPERIMENTAL RESULTS WITH SYNTHETIC 
DATA 
In order to check the accuracy of the algorithm real and 
synthetic data have been used. 
Synthetic data was created using the Microstation SE software 
application, which has the ability to create a camera scene of an 
artificially generated 3D object constructed as a CAD model 
(fig. 2a, 2b and 2c). The 3D object consists of a cube (size of 
1x1x1m) whose sides are marked with crossed red lines. The 
determination of the exterior orientation of the images has been 
calculated using 6 control points with 6, 8um (or 0.8 pixels) 
and RMS error on control points 2cm (pixel size is about 
12cm). 
Six points were measured on the cube using three 
methodologies 
e the least squares image matching technique (LSM) 
e the defocusing algorithm. 
e cross correlation 
The points have been calculated using forward resection and 
the discrepancies of their ground coordinates were compared. 
  
Fig. 3. Gaussian noise was added to the images. The amount of noise was 10% of the image information 
-76— 
LS. 
6.2 
The 
resul 
Corre 
expe 
The 
defo« 
Cross 
from 
techr 
com] 
  
  
Mean 
Stanc 
devia 
(pixe