. Part B3. Istanbul 2004 
ORRESPONDENCE 
N SYSTEMS 
linearly calculated from 
ct linear transformation 
it they must not be in 
of the four points are 
yrrespondences available 
ized. 
) 
where ¢ is an algebraic 
ieous 2-d error provided 
| proper normalization 
this process being a least 
> inclusion of outliers in 
estimation method is 
eliminate the false 
ninimize the influence of 
1e iterative re-weighting 
his method avoids hard 
its. However, the most 
known random sample 
ler; Bolles, 1981). 
thod: Let C=/c,,.,€n} be 
ed from the images. It is 
ion of a set of correct 
ie goal is to identify these 
ize the goal function 
HC y) 
ogenous matrix H while 
linear computation using 
roper subset C, poses an 
The RANSAC-proposal 
X C by drawing random 
druples s=/i, 14} from 
to a hypothesis for H, (at 
for every such hypothesis 
is determined. 
3 
PY 
nsus set of the sample is 
' largest consensus set is 
n for C,. Usually it is too 
nples. There are decision 
its on how many samples 
outlier-rate, a variance for 
dences and a significance 
) continue probing until a 
ched, or — in an any-time 
handed by exterior time 
d in the image the method 
eight on densely populated 
ect correspondence in an 
yn may either end up as 
ght like any other single 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3 
Good Sample Consensus: Already in (Fischler; Bolles, 1972) 
an improvement of the RANSAC paradigm by replacing the 
random samples by samples that are drawn according to an 
assessment criterion is sketched. Following this idea we 
implemented the following approach: 
1. Locations are picked from each image which contain enough 
structure to allow a correspondence test with high significance 
(Foerstner, 1994). More significant locations gain higher 
priority. 
2. Each sample c; is evaluated according to its correlation. 
Samples with high evaluation gain high priority. 
3. Pairs of correspondences (c;,c;) are formed and assessed 
according to their Euclidian distance. Correspondences that are 
far apart gain high priority. 
4. Two pairs form a quadruple (s;,…,s,) of correspondences. It 
is assessed according to the area covered by the smallest of the 
four triangles formed by the points in one of the images. This 
property will be zero if three of the points are collinear. A quad 
with large minimal area gets high priority. 
5. Each quadruple defines a homography using DLT. In the 
space of homographies a metric is defined and quadruples that 
vote for close transforms are merged. The parameters of such a 
cluster of homographies are recalculated using the version of 
DLT that minimizes the residual error R for all correspondences 
preceding it. We call this correspondence set the consensus of 
the cluster. It is assessed according to the size of the consensus 
and also to the assessments of its members and according to 
geometric properties like the size of its convex hull. 
Good Sample Consensus method has been motivated and 
discussed in (Michaelsen; Stilla, 2003). The constructions and 
assessments are coded as productions and entered into a 
production system. The production system is run on the data 
using a data-driven bottom-up control that has any-time 
capabilities (Stilla, 95). 
4. EXPERIMENTS AND CONCLUSION 
4.1 Experiments with Aerial Thermal Videos 
Three video sequences taken from helicopters or aircrafts have 
been used to verify the error behaviour of homography 
estimation and pose estimation using decompositions of such 
homographies. All are taken in the thermal spectral domain. 
Fig. 2 shows example frames for each video. 
Video I: Oblique side-looking sequence on urban region in the 
city of Karlsruhe (buildings on flat terrain) taken with a TICAM 
camera from an airplane flying at approximately 3000m height. 
Such cameras give strong non-projective distortions. The 
camera was zoomed to 540mm focal length. Detector spacing in 
X- and y-direction is 50pm. So this is an extreme tele-lens 
perspective. 
Video II: Oblique side-looking sequence on the same urban 
region (including a lot of homogenous park area) taken with the 
same TICAM camera from an airplane flying at approximately 
3000m height. The camera was zoomed to 212mm focal length 
still giving a small field of view. 
Video III: Forward-looking sequence from a rural region with a 
little creek (trees, bushes, etc.) taken with a focal plane array 
camera from a helicopter flying at very low altitude. Such 
cameras give almost no non-projective distortions. The camera 
has fixed standard field of view. The focal length to detector 
spacing ratio is approximately the same as for Video II. 
    
  
. Istanbul 2004 
  
  
  
  
  
  
  
  
Mi