Roland Geibel
4.2 Irritability by rotation
It would be supposed that a segmentation procedure
produces the same output if the same input data is
presented in different order. Practice shows however that
rotation of the data leads to more or less different outputs
in many of the considered procedures. To test this
dependency the images were turned, segmented and
turned back. Fig. 8b-d shows the results with the input
image turned for 90°, 180° and 270°. The images show
small variations of the contours of small segments and
the appearing or vanishing of small segments. Bigger
segments stay stable. a
. FOM 0 90 180 | 270
rotation
correct 10 11 13 14
over 5 4 5 6
under 4 5 4 2
missed 4 2 1 3
noise 8 10 5 4
q 0.93 0.93 | 0.94 | 0.93
Fig. 8: Rotation of the height image
9. ; S ;
Table 2: Number of classified segments with rotation a) 0°, b) 90°, c) 180°, d) 270°
of the image
By a rotation of the same image data for 90, 180 or 270 degrees each of the investigated procedures showed more ore
less changes in the segmentation. For the procedure Burns the rotation had the least effect.
5 DISCUSSION
During the experiments there arose difficulties in the determination of optimal parameters. In the procedures USF and
WSU this was caused by the big number of parameters and their interdependencies. For this reason it could not be
ensured that the used parameter set was optimal. Some of the investigated procedures assume different sensor specific
recording geometries which however were not present in the data used. In the procedure UB the assumed line-like
sensing order lead to a stripe-like over-segmentation. The own procedure produced the best results. However it also
us oe
ve 4
qe
S ANNE.
Fig. 9: Segmentation results of the FOM procedure with variation of the parameter dy,
a) section of aerial image, b) height image, c) di,-0.45 m, d) diz1.0 m, e) diz1.5 m, f) d-2.5 m
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 333