Estimating 3D Polyhedral Building Models by Registering Aerial Images - Advanced Concepts for Intelligent Vision Systems - page 244

8

1

Average
Maximum

Average
Maximum

7

0.8

6

5

0.6

4

0.4

3

2

0.2

1

0 0

0 10 20 30 40 50 60 0 10 20 30 40 50 60

Noise amplitude Noise amplitude

(a) (b)

Fig. 4. 3D errors (a) and orientation errors (b) as a function of the noise amplitude

show the 3D errors (discrepancy between reconstructed and ground-truth 3D
vertices) and the facet orientation error, respectively as a function of the noise
amplitude.

4.2 Real Data

The proposed approach has been tested with a set of calibrated aerial images
depicting a part of the city of Marseille. These data are provided by the French
National Geographical Institute (IGN). The resolution of these aerial images
is 4158 × 4160 pixels. The ratio between the baseline to the camera height is
about 0.18. One pixel corresponds to a 10 cm square at ground level. Figure 5
illustrates the best model obtained at different iterations of the Differential Evo-
lution algorithm. The projection of the model onto the first and second images is
shown in the first and second columns, respectively. The third column illustrates
the associated 3D model. Figure 6 illustrates the estimated model in cases where
buildings are affected by shadows. Despite the presence of significant shadows
the estimated polyhedral models are correct.

4.3 Method Comparison

To get quantitative evaluation we compared our method with a 3D reconstruc-
tion obtained from Digital Elevation Models (DEMs) 1 . Table 1 depicts the 3D
reconstruction results associated with one polyhedral model (only the heights
of three vertices are shown). The first column corresponds to the reconstruction
obtained with a DEM (robust plane fitting), the second column to our approach
adopting the SSD function, and the third column to our approach adopting the
SAD function. The last row shows the average deviation between the estimated
model and the model obtained with the DEM.

1 Although the DEMs are not ground-truth 3D data, we compared our results with
them for validation purposes.