' 
)0.0 
'aries 
0.0 
10 to 
-2.0- 
  
i ' i 
0.0 50.0 100.0 150.0 200.0 
pixel 
Figure 4.3r Roughness for the scale levels o= 8-12, 40-2 1 
In the case of detailed manual analysis the direct comparison of 
the roughness images on the finer scales (figures 4.2) is a really 
severe problem. The job of solving this problem on the coarser 
(figures 4.3) levels is much simpler. Similarity in this generali- 
zation allows to establish corresponding edges. We do not want 
to recall the statements observed in section 3 with the simulated 
image, though they can also be observed in these real images. 
But one characteristic feature which can be seen in figure 4.21 
is not discussed up to now. In the area marked by a circle (cf. 
figure 4.21) a spurious edge is generated (at x= 26, in level o- 
3.0). But this edge is only present on a small scale range (mo- 
ves to x= 28 in level o- 4.2). 
The contours of the estimated edge locations in scale space for 
this stereo pair are drawn in figures 4.41 and 4.4r. The corre- 
sponding edges are identified and marked. In figure 4.4r two 
dominant edges are present. Unfortunately the left of both is 
just outside the region shown in figure 4.41. 
  
OQ 13.0- 000000000000000000 e00000000000008 Xeenecooseccos 
12.0- s»»e»o5o»»0o9o»2022 © I 
§1.0- sorrononnenoononue » 00000000... ¥eoovoososanas 
4ü.- e-eoeoooooosecsooo 020220000208 Keocoasecesssas 
9.0- 0500000000 0n0000000 © Dean Lane . 
B.Ü- e»e»oosooocos (2)-- » eccccoovcoosioconoooso o cc 
7.0- 000.000...‘ CE o sscase OO 
A A esoXassen us 
5.0- 00000 0 ib © onenc]ansse © ooesenadencnos Meoodses a 00 
  
  
  
4.0- 2200 000 eo * csseXKeosoo o cocseoKenosococcoKseosse v» ce 
J0- 2... veau» auskennen 9 wonnovMouv 0 6 a0vvfv0sues ee 
20-000 0 edere conos» c 00000 ses» e © and kenen nenn c 
40-0000 ancdfe © Hevescos 3c» cocco He a soo coder c» Je 0 
0.0 - ==> » codo © dffrocuo » mj o mo Hero 3» ^ 
48.0 58.0 68.0 28.0 88.0 98.0 
pixel 
Figure 4.41 Location of edge candidates in scale space. The de- 
tected edges are marked by * (left stereo partner) 
301 
D0 - ovesssevonve MKeaseesosoeva es9^2  « ssoconeus Xe enooe 
5.0- *eoooosoccinossoceonecesieessn o cccovoceveeeecee 
4ü- nones rr: = soncofeesse » scceces ¥evoonose 
= sme seesfassss © susssKess se a u ecke 
2.0- sms © neun kenn = = 
1.0- =o eo s c run = v6 0 offre 0 swe » uncoukes 0 nu sdf 
0.0; see noo modifs © à vo due © mms eoe) o oor o HK 
48.0 58.0 88.0 78.0 88.0 98. 
pixel 
Figure 4.4r Location of edge candidates in scale space. The de- 
tected edges are marked by * (right stereo partner) 
Finally we address some consequences of tracking edges 
through scale for matching. The correspondence of edges is 
established manually and with signal control. We start (a) the 
tracking at level o= 3 and (b) at level o- 12. The results are 
listed in table 1. 
Table 1 Influences of tracking to point location and matching 
  
image left image right left - right 
AX (pixel) AX (pixel) Apx (pixel) 
mad mse | mad mse | mad mse 
  
° start” ? 
103 0,62 | 0,54 0,77 
N= 15 points 
e start” 12 
1,05 0,55 | 0,31 0,39 | 0,89 0,68 
N= 8 points 
  
  
  
  
  
  
AX measures the dislocation of the edges between the start level 
(ostart) and the level o= 0, i.e. the unsmoothed image. Apx 
measure the difference in these dislocations, i.e. it gives the 
resulting systematic effect for the parallaxes. mad stands for 
mean about differences and mse denotes the mean square error. 
This statistic shows that in theses cases a dislocation of about 1 
pixel is observed on the average. The largest shift between two 
levels (ao= 1) amounts to 2.9 pixels. The area in which this 
shift occurs is marked by the rectangle in figure 4.41. The syste- 
matic effect for the parallax nearly reaches about 1 pixel on the 
average. Because such a parallax error directly propagates on 
the calculated heights, we have to investigate this effect in more 
detail in future.