ISPRS Commission III, Vol.34, Part 3A , Photogrammetric Computer Vision*, Graz, 2002
Figure 2: Two of the 500 images that were used to test the robust-
ness against occlusions.
Figure 3: Two of the 200 images that were used to test the robust-
ness against arbitrary illumination changes.
objects that occlude the IC generate clutter edges, this actually
underestimates the occlusion.
The transformation class was restricted to translations, to reduce
the time required to execute the experiment. However, the al-
lowable range of the translation parameters was not restricted,
i.e., the object is searched in the entire image. Different values
for the parameter of the minimum score were applied, which can
be chosen for all approaches. The minimum score specifies the
score a match must at least have to be interpreted as a found ob-
ject instance. The forward fraction of the Hausdorff distance was
interpreted as score value. Initial tests with the forward and re-
verse fractions set to 30% resulted in run times of more than three
hours per image. Therefore, the reverse fraction was set to 50%
and the forward fraction was successively increased from 50% to
90% using an increment of 10%. The parameter for the maxi-
mum forward and reverse distance were set to 1. For the other
three approaches the minimum score was varied from 10 to 90
percent.
To test the robustness against arbitrary illumination changes, a
second sequence of images of the IC was taken, which includes
various illumination situations. Two example images are dis-
played in Figure 3. Due to a smaller distance between the IC and
the camera, the ROI is now 255 x 140 pixels containing 3381
model points on the lowest pyramid level. The parameter settings
for the six methods is equivalent to the settings for testing the
robustness against occlusions.
3.2.2 Accuracy In this section the experimental set-up that
we used to determine the accuracy of the algorithms, is explained.
This criterion is not applied to the Hausdorff distance, because
subpixel accuracy is not achieved by the used implementation.
Generally, it seems to be very difficult to compute a refinement of
the returned parameters directly based on the forward or reverse
fraction. Since PatMax® and the shape-based matching are the
only candidates that are able to recognize scaled objects, only
the position and orientation accuracy of the five approaches were
tested.
To test the accuracy, the IC was mounted onto a table that can be
shifted with an accuracy of 1 um and can be rotated with an accu-
racy of 0.7° (0.011667°). Three image sequences were acquired:
In the first sequence, the IC was shifted in 10 um increments
to the left in the horizontal direction, which resulted in shifts of
about 1/7 pixel in the image. A total of 40 shifts were performed,
while 10 images were taken for each position of the object. The
IC was not occluded in this experiment and the illumination was
not changed. In the second sequence, the IC was shifted in the
vertical direction with upward movement in the same way. How-
ever, a total of 50 shifts were performed. The intention of the
third sequence was to test the accuracy of the returned object ori-
entation. For this purpose, the IC was rotated 50 times for a total
of 5.83°. Again, 10 images were taken in every orientation.
During all accuracy tests, euclidean motion was used as transfor-
mation class. The search angle for all approaches was restricted
to the range of [-30°;+30°], whereas the range of translation pa-
rameters again was not restricted. The increment of the quantized
orientation step was set to 1°, which results in the models con-
taining 61 rotated instances of the template image at the lowest
pyramid level. Since no occlusions were present the threshold
for the minimum score could be uniformly set to 80% for all ap-
proaches.
3.2.3 Computational Time In order to apply the third crite-
rion, exactly the same configuration was employed as it was used
for the accuracy test described in Section 3.2.2. The computa-
tion time of the recognition processes was measured on a 400
MHz Pentium II for each image of the three sequences and for
each recognition method (excluding again the Hausdorff distance
for the reason mentioned above). In order to assess the corre-
lation between restriction of parameter space and computation
time, additionally, a second run was performed without restrict-
ing the angle interval.
In this context it should be noted that the modified Hough trans-
form is the only candidate that is able to recognize the object,
even if it partially lies outside the search image. The translation
range of the other approaches is restricted automatically to the
positions at which the object lies completely in the search image.
Particularly in the case of large objects this results in an unfair
comparison between the Hough transform and the other candi-
dates when computation time is considered.
3.3 Results
In this section we present the results of the experiments described
in Section 3.2. Several plots illustrate the performance of the
examined recognition methods. The description and the analysis
of the plots are structured as in the previous section, i.e., first
the results of the robustness, then the accuracy, and finally the
computation time are presented.
3.3.1 Robustness
Occlusion. First, the sequence of the occluded IC was tested.
A complete comparison of all approaches concerning the robust-
ness against occlusion is shown in Figure 4. In the left plot the
recognition rate, which is an indicator for the robustness, is plot-
ted depending on the minimum score (see Section 3.2.1). Here,
the superiority of our two novel approaches to the standard ap-
proaches becomes clear. Note that the robustness of the modi-
fied Hough transform hardly differs from the robustness achieved
by the shape-based matching. Looking at the other approaches,
only PatMax® reaches a comparable result, which is, however,
slightly inferior in all cases. Furthermore, when using a restricted
parameter space, which is limited to only translations as described
in Section 3.2, the recognition rate of PatMax® was up to 14%
lower as when using a narrow angle tolerance interval of [-5°;+5°].
A - 371
