International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B3, 2012 
XXII ISPRS Congress, 25 August - 01 September 2012, Melbourne, Australia 
  
0 10 20 30 40 S0 60 70 8 90 100 0 10 20 30 40 S0 60 70 80 90 100 0 10 20 30 40 S0 60 70 8 90 10 0 10 220 30 4 S0 6 70 80 90 100 
(a) CT, Cones (b) CT, Teddy (c) CT, Venus (d) CT, Tsukuba 
     
% RAC 8 BRD FE 
    
  
0 888888888888888 
(e) CT, P. (f) CT, Pj (b) CT, P», 
  
  
   
Teddy 
    
— Venus Tsukuba -Cones ^^ Teddy — Venus «+ Tsukuba 
      
   
0 100 200 300 400 500 wo » Ko “a0 1000 1100 120 1x0 1400 1500 0 100 200 200 400 500 P3 "0 800 "500 ^ 1000 100 1x0 1300 1400 1500 
(1) Ordered Configurations for P»; (j) Ordered Configurations for P»; 
15.0% 
14.0% 
1309 
120% oo 
11.0% 
  
10.0% 
  
   
0 10 20 30 40 SO 60 70 80 90 10 0 10 20 30 40 50 6 70 80 90 100 0 10 20 30 40 50 6&0 70 80 90 100 
(k) RT, Cones (1) RT, Teddy (m) RT, Venus (n) RT, Tsukuba 
Figure 4: Errors in unoccluded areas for all penalty functions employing census transform examining behavior on the same image (row 
one), over different images (row two), stability of configurations across images (row three), and employing rank transform (row four). 
   
  
ares LU 
  
(a) Left Cones image (b) Ground thruth (OCT Re (d) CT, P», (e) CT, BP; (CT. 
  
     
E 
(g) Left Teddy image (h) Ground Truth (1) CE P. Q CT, P1 (k) CT, P», (D CT, P», 
Figure 5: Disparity maps obtained with optimally parametrized penalty functions for Cones (top row) and Teddy (bottom row). 
than for the census transform which is due to the missing spatial comparison, these two functions are plotted with their optimal 
information of the rank transform. As with the census transform configuration in Fig. 6 showing an obvious similarity between 
P», performs acceptably and P», poorly for an adaptive func- the two functions over x. 
tion. However, in opposite to census, P», always outperforms 
Py; in 3 cases quite significantly. All functions are similarly 3.2 Simulated Degenerated Images 
robust towards parametrization offset as with census (data not 
shown). Again, good configurations coincide across all images For the second set of experiments the left input image of the 
(data not shown). Cones data set is artificially degraded whereas the right remains 
unchanged. Two types of radiometric differences and two types 
The error levels obtained with optimally adapted penalty func- of noise are considered: 
tions are summarized in Table 1. For the two test images Cones 
and Teddy the resulting disparity maps are shown in Fig. 5. Even e Additive white Gaussian noise (AWGN) with SNR = 12 dB 
visually highly noticeable, P», introduces a significant amount of e Salt-and-Pepper noise with 14 96 of the image degenerated 
errors. Using P5; the small structures in Cones are retained, oth- e Linear brightness (gain) change across the half the image 
erwise there is no significant difference between P», and P»;. For e Non-linear brightness (gamma) change 
Figi 
ima, 
The 
uati 
exp 
the 
age 
pare 
func 
ima 
gen 
quic 
to tl 
are 
non 
case 
Pep] 
cept 
forn 
out[ 
eval 
Con 
atio 
As | 
ofte 
evel 
non 
and 
eter 
P 
disp 
£y — 
Po 
in € 
mal 
in F 
pen 
leve 
ini 
tion 
Pre] 
3.3 
For 
tren 
opti 
Was 
deg 
paic 
(e. ; 
usir 
acci