———————— S 
10.9 mm on the 
it is essential to 
its of the image, 
ollowing section 
ages. The points 
en images and à 
r, 1988; Heipke, 
ost sophisticated 
on an image the 
due to repeated 
ry stations. The 
Menci Software 
age matching a 
e Locus; Kraus, 
eoscopic images 
Xtervention. The 
ree-dimensional 
… The automatic 
] in the working 
it entered by the 
which is easily 
(orthogonal). 
iting before and 
dpi resolution, 
iced during the 
oftware. 
simultaneously, 
how an average 
Bitelli, Gabrielle 
The area tested, measuring approximately 10 m? is the lower right-hand part of the structure. It is reproduced in two 
photograms (bottom in figure 2) and, therefore, a single photogrammetric model. 
7 DSMs were automatically generated, on a square grid with grid interval 5 cm, changing some matching parameters 
each time. To start the matching, the same grid of 334 basic points manually acquired at the digital station in 
stereoscopic mode was employed, with grid interval of 50 cm. In order to validate the models generated in this way, a 
reference DSM of 8428 points was created using the Digicart 40 analytical plotter, with the same coordinates of the 
origin and grid interval 10 cm. 
Comparisons were made by subtracting the models generated using StereoView point-by-point from the reference 
DSM. The initial tests were carried out by varying the half-width of the search matrix (AX = AY = 7, 9, 11, 15, 21, 31 
pixels) and repeating automatic matching, starting with the same basic information (same model, orientations and set of 
basic points). The half-width is related to the horizontal and vertical dimension of the search matrix, expressed in 
pixels: a value of AX=AY=10 pixels defines an area of size 20x20 pixels on each image. The matrix size, expressed in 
the object reference system, can be calculated by multiplying the size of the search window in pixels by 4 mm (pixel 
size in the object system). 
The results of the comparisons and some statistical parameters for the differences between the model produced by the 
analytical plotter (D) and those generated automatically (7+31 spot pixels), are indicated in the following table. 
  
  
  
Search DSM percentage frequency of point differences 
window size points |Mean|Std. Dev.| Min/max 
Image |Object| (n.and |(mm)| (mm) (mm) |-400 + -10|-10 +-5|-5 + -2/-2 + 2/2 + S[S + 10[10 + 200 
(pixels)| (cm) |96 Cc20.7) cm cm cm | cm | cm | cm cm 
D-7 14 5.6 | 4748 (56) | -23 230 |-3479/1554| 15.50 8.80 | 10.80 |30.00|16.00|10.00| 8.90 
D-9 18 7.2 | 4872 (58) | -14 180 |-3479/1738| 11.40 7.50 | 11.80 |36.90|17.70| 8.40 | 6.30 
D-11| 22 8.8 | 5015(60) | -14 138 |-2171/1653| 10.60 6.70 | 12.40 |40.40|17.60| 7.50 | 4.80 
D-15| 30 12.0 1 5277 (63) ] -11 127 |-2035/1669| 9.50 6.30 | 11.10 |45.40|17.30| 6.60 | 3.80 
D-21| 42 16.8 | 5513 (65) | -17 133 ]-2013/1732| 9.50 6.50 | 11.30 |49.00|16.20| 4.40 | 3.10 
D-31| 62 24.8 | 5547 (66) | -19 132 |-1095/1622| 12.2 9.3 9.0 |45.90| 14.0 | 4.9 4.7 
Table 2 — Assessment of the differences between the reference model generated with the analytical plotter (D) and the 
seven surface models (spot size 7+31 pixels) generated by matching procedure: search window size, number and 
percentage of points with correlation coefficient 20.7, main statistical parameters, percentage frequency of point 
differences. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
The point comparison between the DSM created using the Digicart 40 analytical plotter and the DSMs produced 
digitally using the StereoView software, allowed the following conclusions to be drawn: 
- the table indicates that the percentage of generated points, a function of the matching coefficient value used, is always 
lower than 70% and rises in direct proportion to the size of the search window; 
- the residuals divided into classes shows that of the various tests carried out, the best solution is that in which the 
search window measures 16.8 cm (spot size — 21 pixels). In this solution, the frequency of residuals which belong to the 
central class of +2 cm is the highest and the percentage of residuals relative to the extreme value classes is at a 
minimum. The solution is worsened by bringing the search window to 24.8 cm (spot size = 31 pixels); 
- the statistical parameters for the depth differences, such as standard deviation, reveal that, in contrast to the previous 
conclusion, the best solution seems to be that obtained using a 12 cm (spot size = 15 pixels) search window. Said 
statistical parameters, calculated on the entire population of depth differences, are not sufficient to validate a DSM, 
since this requires a spatial evaluation of their trend on the object to identify local deficiencies on the model. It must be 
emphasised that the extreme incorrect values (outliers) have a significant influence on the depth difference statistics. 
Therefore, it is sufficient for the stereoscopic model to have occluded or shadow zones to automatically generate points 
with large errors. 
The tests carried out revealed how the size of the image block, that is to say, the size of the search window, is a 
fundamental parameter in the image matching effected with the software. 
Further tests, localised in the balcony area and, therefore, with the greatest depth difference (approx. 2.5 m), showed 
that the spot size of 21 pixels used was optimum for automatic imaging of the entire facade, both for flat zones and for 
zones with sudden depth changes. 
The depth difference graphs (figure 3) show that on the flat areas of the object, increasing the search window visibly 
increases the points plotted correctly. The “salt and pepper" appearance is reduced, increasing the areas of homogenous 
distribution and the solution becomes more stable overall. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 65