Roland Geibel
q=( Ys * Yos: + S ku Su. )/|T A| (13)
TeTK TeTo MeMy
For the over-segmented and for the under-segmented segments the ratio of intersection S ro and Spy are taken into
account as weight factors too. However their contribution is reduced by the additional factors ko «1 and ky<1. The
influence of over-segmented and under-segmented segments should be different. For an over-segmentation it is possible
that in one of the following processing steps M-segments are combined if they belong together. For an under-
segmentation a subsequent splitting based on the features of the fused segment may be difficult. For visualisation tasks
also the effect of an over-segmentation is not as bad as that of an under-segmentation if the normal vectors of the
segmented parts were determined correctly. As weight factors for ko and ky; we propose:
ko z 2n-1) / n? with n 2 number of M-segments associated with the T-segment (14)
and
ky = 1/ m? with m = number of T-segments associated with the M-segment. (15)
The segments classified as missed worsen the quality q because they make no contribution. Segments classified as noise
worsen the result and therefore should be considered in reducing the quality. For example the sum in the numerator of
equation (12) can be decreased for the number My| of missed segments. However since the size of the segments was
not yet recognised a few small noise segments would have much more effect than one big segment. In general also the
size of segments should be recognised in their importance for the total area. As a measure of quality we propose:
dLo =( YS Ar S. koSr Art Y vau. uy “AT JIM 7 Sin f > Ar (16)
TeT& TeTo MeMy MeMN TeTA
qr 7 max (0,q; o) (17)
For the evaluation of a complete scene the quality number qq has to be derived from the results of all of the segments
of the scene or from the results qj, of the objects. For the evaluation of a scene we use
Qc 7 Ya Ar =, Aj, with A, = area of the local object (18)
local objects local objects
4 EXPERIMENTS
4.1 Quantitative evaluation of the procedures by an example
In order to test and evaluate the segmentation
procedures laser altimeter data of real objects in Number of | Burns | USF | UB | WSU | FOM
the area of the town of Karlsruhe were used. The segments
resolution of the data is 1 m^ on ground and the correct 8 7 1 5 10
resolution in altitude is 0.06 m. The scenes were over 1 3 10 3 5
recorded in "first-pulse-mode". For adjusting the under 5 0 3 3 4
segmentation procedures many experiments were missed 2 17 9 2 4
made for each of the procedures. The parameter noise 7 9 4 5 8
adjustment was done for one selected object of q 0.33 0.54 | 0.38 0.23 0.93
the test area. The results of the investigated
procedures with the parameters valued as best are Table 1: Number of classified segments by different
depicted in Fig. 1 and 2d. The evaluation of the procedures
results is shown in table 1. The quality measure
qu was determined using the weight factors ky
and ko proposed in equation (14) and (15).
332 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
