Gruen, Armin
Figure 4. House extraction with DIPAD. Several stages (a-d) of the iterative refinement procedure (after Streilein, 1999)
An issue which deserves more attention in the future is texture modeling. As the quality of models improves over time
more emphasis has to be put on correct texture rendering. At least for visualization purposes there exists a strong duality
between texture model and geometry model. Insufficient modeling of geometry can easily be obscured by realistic
texture. Videogame developers are fully aware of this fact and use it extensively in order to reduce the number of
polygons in object modeling for real-time rendering. A fine approach to view-dependent texture mapping has been
suggested by Debevec et al., 1996.
4 SITE MODELING AT IGP, ETH ZURICH
In recent years we have developed a number of techniques for automatic and semi-automatic site reconstruction and
modeling (see chapter 3 and Streilein, 1994, Henricsson et al, 1996, Gruen, Li, 1997, Gruen, Dan, 1997, Gruen, Wang,
1998, Niederoest, 2000, Zhang, Baltsavias, 2000, Park, Zimmermann, 2000, Zimmermann, 2000). Here we will report
about two new approaches, which are aiming at operational (precise and reliable) extraction of objects and are as such
of special relevance to the professional practice.
4.1 ATOMI
ATOMI (Automated reconstruction of Topographic Objects from aerial images using vectorized Map Information) is a
project, supported by the Swiss Federal Office of Topography, Bern, to develop and test algorithms and develop
operational software for the updating of road centerlines and building outlines of given digital map data 1:25 000. In
addition, for building roofs one representative height has to be assigned. The map vector data is used as approximation,
guiding the image analysis algorithms. The processing should be as automatic as possible, but it is accepted that some
cases need manual guidance and editing (Eidenbenz et al., 2000).
For building extraction and updating the system uses multiple cues and works with color aerial images. A Digital
Surface Model (DSM) and an orthoimage are automatically derived from a stereomodel. Blob detection in the DSM and
unsupervised multichannel classification are used to produce approximate vector data. This allows to introduce
buildings into the dataset that are not yet in the existing vector map. Multichannel classification utilizes the following
information channels:
+ Shadows, derived from the S-channel of the HIS color space by thresholding
Channel a* from the CIELAB color space
Texture channel (texture measure is the number of edge pixels within a circle around a center pixel)
DoA (Degree of Artificiality) channel, with DoA derived from (G-R)/(G+R). DoA separates man-made structures
from vegetation (Sibiryakov, 1996)
Channel containing the normalized DSM (nDSM = actual DSM-DTM)
+++
+
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 313