ng
on
ler
on
ed
1at
re.
ch
en
lar
Sagi Filin
3 EXPERIMENTAL RESULTS
The algorithm was applied to several data sets each composed of a vector compilation of a 1:50,000 scale map and from
data (road network) digitized from respective orthophoto maps. The difference between the two sources was great in almost
every aspect. First the amount of detail in the 1:50,000 maps was far greater than in the other set, then the period of time
between the compilation of the both sets generated major changes in content. The average positional difference between the
data sets varied between 70m and more then one hundred meters (depending on the data set that was used), the maximal
distance we encountered in several data sets was in the order of 300m. Figure 7 illustrates some of the aspects mentioned
above.
The difference has shifted our concern in evaluating the
algorithm to two aspects, the improvement in accuracy
a A. achieved by applying the algorithm and the level of
un i MA j^ ^ automation achieved in the counterpart detection
i MÀ a A UE Te algorithm. The second concern had mostly to do with
hu "uo A enu. eds , An assessing the fitness of the counterpart detection algorithm
TI E oor and the level of success in automation of this process. Let
"Hy us first address the matching procedure. The number of
nodes (the matching was based on detection of counterpart
nodes between the data sets) received from the GIS data
was usually in the upper thousands while the order of
nodes from the photogrammetric based GIS was usually
half of that. The amount of counterparts that were detected
was usually 50% or more of the number of nodes in the photogrammetric data. The percentage of correct counterparts
detected by the algorithm was higher than 90% (the lowest percentage was 92.5% and the highest 96.4%). However our
major concern was alerting about the presence of suspicious counterparts. An evaluation algorithm we developed for
reliability of counterparts enabled visiting and correcting suspicious counterparts. The amount of nodes that were needed to
be visited was mostly on the order of 10% (of which ~ 40% were wrongly classified). These nodes tended to cluster around
areas with poor correspondence and usually by correcting few, many others were corrected as well as a follow-up by the
application.
Figure 7. An example for differences between data sets
To evaluate the improvement in positional accuracy we evaluated the improvement in the accuracy of roads that did not take
part in the transformation (mostly secondary roads). These secondary roads were digitized from the orthophoto maps and
served as a “comparison” sample set. The coordinates of these features (as digitized from the 1:50,000 topographic maps)
were corrected according to the presented algorithm, and the corrected coordinates were checked against the "comparison"
sample set. It was found the improvement in accuracy that was on the average on the order of 75%, so that the average
distance between the objects original location and their true location was reduced, for example, from 44m to 9m on average,
113m to 30m and so on, whereas the MSE values were reduced similarly from 49m to 11m, 114m to 34m and so on.
Evaluation of other GIS objects such as position of creeks and others objects, have shown similar results.
4 SUMMARY
Utilizing GIS data to improve object recognition tasks has not received the attention it deserves, most likely due to its not
being part of the vision paradigm. However, since it exists and since the main task of object recognition from aerial imagery
is the creating and updating of GIS databases, it is only reasonable to use the data already in existence. In this paper, we
have addressed a preliminary task that is usually overlooked, concerning registration of one data set to another. Indeed this
task was preformed in the past by different methods, for example using global transformations or using point based local
transformations. However, issues such as having counterpart features coalesce or improvement in accuracy of related
objects were not of much concern. The current algorithm addresses both these issues and performs this task in a more
natural way in respect to the data in question.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 287