Previous : Algorithm Up : Contents Next : Conclusion



Results

We present here a result obtained on a part of the DEM of the city of Amiens (France). Here, small intersections are allowed, for 3D visualization reasons.

The proposed algorithm gives a list of buildings in a text file. Figure l) shows the result of the detection a few comments can be made :
  • there are some border effects : in the top left corner of a building is irrelevant. This is due to the ground estimate, which is made with only 2 points, since other points of the upper side are out of the area. These kind of phenomena can be regulated by imposing a minimum number of points while estimating values like ground height.
  • some small and low buildings are missed, because of the structure of the data which is too smooth where buildings are low.
Figure m) shows the results of roof estimation (top) and the ground truth (bottom) provided by the French Mapping Institute (IGN). This ground truth is precise since it has been built by hand using very high resolution aerial images. Some buildings are missing, and details of the shape of buildings are missed, due to the rectangular silhouette model we have chosen.

Figure n) shows the data and the result in 3D, and allows to see that mistakes are due to artifacts on the data.

Comments

The results presented here have been obtained in 40 minutes with a SUN-blade 2 (500 MHz, 250 MB). The data image size is 1060 by 1024. We should be able to deal with even worst data, if we could optimize more complex energies. Tests have been done (see [14]) with more complex prior models, using relations such as alignment or orthogonality between buildings.

Results are not improved with a more complex prior model due to the optimization step. Simulated annealing looses its efficiency if the energy exhibits too many local minimas. Designing a good prior model and estimating its parameters is then hard to achieve, since we cannot obtain optimal configurations.

Therefore, one way to improve the proposed technique is to work on a more efficient optimization algorithm. This will allow to improve the model and use more complex silhouettes and a bigger set of possible roofs. This will be done in a near future.

Figure l) Original crude DEM and result of the building extraction. (click on the image to get the entire scene).




Figure m) Estimated roofs and ground truth manually obtained by IGN.




Figure n) Results in 3 dimension. Top : original crude DEM, bottom : refined DEM obtained by the proposed method.





Previous : Algorithm Up : Contents Next : Conclusion



Back to M. Ortner's home page !