In: Paparoditis N., Pierrot-Deseilligny M„ Mallet C., Tournaire O. (Eds), IAPRS, Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
5.2 Experiments on real data 
In order to compare our algorithm with PROSAC on real data 
we have collected a test set that consists of 28 pairs of 
overlapping aerial images. Those images have been taken with 
different cameras under varying conditions and include both 
digital and analog photos. The set also provides a reasonable 
variety of areas presented in the images as they include forests, 
fields, towns, mountains and rivers. All the images have been 
downsampled to the size of ~0.5 mega pixels. The overlap 
percentage is less than 10% for all the pairs, while the rotation 
angle varies from 0 to 180 degrees. 
To provide both our algorithm and PROSAC with putative 
matches we first detected 1000 Harris corners (Harris et al., 
1988) per image which is enough to cover the whole image area 
uniformly. SIFT descriptor with radius r = 12 pixels was then 
applied to those feature points and, finally, a simple nearest 
neighbor matching procedure was carried out. The descriptors 
distance threshold was set to -J80000 , as recommended in 
(Labe et al., 2006). Matches for PROSAC were sorted by the 
nearest neighbor distance ratio. 
(b) 
Figure 5. Results of the proposed method on real data. Overlap 
area is 3.5% in (a) and 10% in (b) 
Like in synthetic tests, our algorithm showed an advantage over 
the PROSAC-based scheme when applied to real data. Being 
absolutely deterministic, as opposed to SAC-methods, the 
proposed algorithm demonstrated better robustness and 
successfully matched 71% of the challenging test pairs, while 
PROSAC coped with 54%, often having to run through 
thousands of iterations to find a solution. The results are shown 
in Figure 5. 
In our experiments we also tried an alternative approach to 
rotation angle detection. As we use SIFT descriptor, the 
dominant direction of each feature is known. Thus, it is possible 
to vote for the difference in directions by each match pair to 
determine the relative orientation of the images. But this 
approach turned out to be less robust than the one described in 
Section 4.1, due to instability of feature direction detection by 
SIFT and much lower total amount of votes. Moreover, this 
approach depends on a specific descriptor, while the proposed 
method is able to work with pure point correspondences. 
5.3 Matching a set of images 
The algorithm of shift-rotation model estimation has been used 
in a framework that matches a whole set of aerial images. The 
amount of images in a set is -100-200. A set is usually divided 
into a number of routes. For each image the route to which it 
belongs is known beforehand and no other additional 
information is used. Note that, while the overlap among images 
belonging to the same route is substantial, typical overlap 
between neighbouring routes is on the order of 20-30% or even 
less in the most complex cases. Moreover, in general case the 
routes may be placed randomly with respect to each other. Thus, 
if no GPS / IMU data is available, pairs of overlapping images 
must be extracted automatically. We used the approach 
proposed in (Brown et al., 2007) for candidates’ selection and 
then applied our method for image matching. The overlapping 
images were successfully detected in most cases which made an 
automatic assembly of image blocks possible. 
6. CONCLUSIONS 
We have presented a new method of shift-rotation model 
estimation that is based on a voting procedure in parameter 
space. The proposed method improves the stage of model 
estimation within the bounds of the discussed two-level scheme 
of matching aerial images. This allows a more robust and 
qualitative detection of overlap area which results in better 
matching of aerial images with low overlap. Our method has 
been shown to outperform the existing approaches, particularly 
the modem PROSAC algorithm. The results of the synthetic 
tests have been proved by experiments on real data that 
consisted of an exhaustive set of aerial images with small 
overlap. The advantages of the proposed method are 
determinacy and a straightforward set of parameters that do not 
require tuning. 
REFERENCES 
Begg, C., Mukundan, R„ 2004. Hierarchical Matching 
Techniques for Automatic Image Mosaicing. In: Proceedings of 
the Conference on Image and Vision Computing, Akaroa, New 
Zealand, pp. 381-386