CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
The strength of this process is its ability to localize accurate 
global structure breaks: it separates façades and foreground. On 
the one hand, split results at the foreground are not really inter 
esting because the related region is not in the rectified plane: they 
are based on chaotic gradient distribution. In such a case, the 
process stops or it oversegments. This phenomenon typically oc 
curs on the cars of figure 11. On the other hand, splits inside 
façade texture provides some significative information. On fig 
ure 10, the left façade is first split between the second and the 
third floor, whereas the first windows column is extracted from 
the right façade. This different strategy certainly must be ex 
plained by the fact that the process is exclusively based on edges 
alignment. An other criterion like contour uniformity may direct 
the split decision toward a more significant separation: favoring 
floor separation rather than window columns. 
Figure 3 shows the region models, the leaves of the segmentation 
tree. One can see that the synthetic image reconstructed from the 
2Z)-models is very close to the initial image although the rep 
resentation is very compact. This shows that our modelling is 
particularly well adapted for image compression. 
Figure 12: Upper: Rectified façade image. Bottom: Synthetic 
image reconstructed from 1000 elementary 2£>-models. 
7 CONCLUSIONS AND FUTURE WORK 
In this paper, we have presented a new unsupervised model-based 
segmentation approach that provides interesting result. It is able 
to separate a façade from its surroundings but also to organize 
façade itself in a hierarchy. Still these are first results, thus there 
are many improvements that could be made. The dictionary of 
models is currently being extended to periodic textures to man 
age for instance balconies, building floors or brick texture. Some 
other objects or specializations of objects could be added such 
as symmetry computation of (Van Gool et al., 2007). A merger 
process at each step of the process could also be useful to correct 
oversegmentations. Besides we could add color information to 
directly detect difference between two façades ore between two 
floors in certain cases. We could also use a point cloud to com 
pute an ortho image: displacements due to perspective effects 
would be avoided. 
Such an unsupervised segmentation will provide of course rele 
vant clues to classify the façade architectural style or to detect 
objects backward or in front of it. It is also intended to give geo 
metrical information that represents relevant indexation features 
e.g. windows gab length or floor delineation. 
REFERENCES 
Alegre, F. and Dellaert, F, 2004. A Probabilistic Approach to the 
Semantic Interpretation of Building Facades. Technical report, 
Georgia Institute of Technology. 
Ali, H., Seifert, C., Jindal, N., Paletta, L. and Paar, G., 2007. 
Window Detection in Facades. In: Proc. of the 14th International 
Conference on Image Analysis and Processing, pp. 837-842. 
Han, F. and Zhu, S.-C., 2005. Bottom-up/top-down image pars 
ing by attribute graph grammar. In: International Conference on 
Computer Vision, IEEE Computer Society, pp. 1778-1785. 
Kalantari, M., Jung, F., Paparoditis, N. and Guédon, J.-P., 2008. 
Robust and Automatic Vanishing Points Detection with their Un 
certainties from a Single Uncalibrated Image, by Planes Extrac 
tion on the Unit Sphere. In: IAPRS, Vol. 37 (Part 3A), Beijing, 
China. 
Korah, T. and Rasmussen, C., 2007. 2D Lattice Extraction from 
Structured Environments. In: International Conference on Image 
Analysis and Recognition, Vol. 2, pp. 61-64. 
Lee, S. C. and Nevatia, R., 2004. Extraction and Integration of 
Window in a 3D Building Model from Ground View images. In: 
Computer Society Conference on Computer Vision and Pattern 
Recognition, Vol. 2, pp. 113-120. 
Miiller, P., Zeng, G., Wonka, P. and Van Gool, L., 2007. Image- 
based Procedural Modeling of Facades. In: Proceedings of ACM 
SIGGRAPH 2007 / ACM Transactions on Graphics, Vol. 26num- 
ber 3, p. 85. 
Reznik, S. and Mayer, H., 2007. Implicit Shape Models, Model 
Selection, and Plane Sweeping for 3D Facade Interpretation. In: 
Photogrammetric Image Analysis, p. 173. 
Ripperda, N., 2008. Determination of Facade Attributes for Fa 
cade Reconstruction. In: Proc. of the 21st Congress of the Inter 
national Society for Photogrammetry and Remote Sensing. 
Cech, J. and Sara, R., 2007. Windowpane Detection based on 
Maximum Aposteriori Probability Labeling. Technical report, 
K13133 FEE Czech Technical University, Prague, Czech Repub 
lic. 
Van Gool, L. J., Zeng, G., Van den Borre, F. and Müller, P., 2007. 
Towards Mass-Produced Building Models. In: Photogrammetric 
Image Analysis, p. 209. 
Wenzel, S. and Förstner, W., 2008. Semi-supervised Incremental 
Learning of Hierarchical Appearance Models. In: Proc. of the 
21st Congress of the International Society for Photogrammetry 
and Remote Sensing.