primitive objects is approximately the same across the image. A test was also
performed with an increased scale parameter (Figure 9.4f). The developed edgeconstrained segmentation performed even better at larger scales, while the stripping
problem was less apparent (Figure 9.4f).
For the QuickBird imagery (Figure 9.5a) a similar test was performed and
demonstrated (Figures 9.5e,h,f,i) within a semiurban area. The developed algorithm
FIGURE 9.5 Comparison of various segmentation algorithms on QuickBird data set (Eastern
Attika, Greece): (a) original image; (b) mean-shift segmentation with default parameters;
(c) watershed segmentation with default parameters; (d) multiresolution segmentation (eCognition) with default parameters; (e) canny edge detection applied on AML scale-space representation; (f) LSD line features extracted from original image; (g) standard MSEG results with scale
parameter 100; (h) edge-constrained MSEG with Canny edge features used and scale parameter
100; (i) edge-constrained MSEG with LSD line features used and scale parameter 100.
EVALUATION AND DISCUSSION
189