In this work we introduced methods to allow semantic based interaction in highly detailed 3d terrain- and city models. In the proposed approach different ontologies can be used without changing the geometric LOD-hierarchy used for photorealistic rendering. In our opinion, the ability to switch the ontologies and thus the categories in which we think about the data is more important than the exact correspondence between geometric features and semantic entities. We have presented methods which allow interaction with photorealistic visualizations, so that the missing explicit link becomes unnoticable to the user.
6.5.1 Future Work
The ability to access detailed semantic information within a real-time photorealistic rendering framework is just a first step towards semantic interaction. With the combined strength of detailed semantic models and visually detailed instances there is a lot of place for new interaction paradigms especially concerning the modeling or synthetization of such scenes but also the exploration and visualization techniques.
Another direction of research will concern the flexibility of models. By now, all successful realtime rendering methods have to preprocess the data. Apart from the delay which is introduced by doing so, there is a conceptual difference whether the data needs to be static or it can be adapted on-the-fly.
6.5. CONCLUSION
Figure 6.3: Screenshot from the rendering application. In picking mode semantic entities are highlighted as the mouse hovers over them in the image. The user thus gets an instant feedback with which objects and on which semantic LOD he is about to interact.
Acknowledgments
This work was funded by the German Science Foundation (DFG) as part of the bundle project “Abstraction of Geographic Information within the Multi-Scale Acquisition, Administration, Analysis and Visualization”. We thank the Berlin Senate and Berlin Partner GmbH for making the CityGML model available. Finally, we also want to thank the reviewers for helpful comments.
CHAPTER 7
C ONCLUSION AND F UTURE D IRECTIONS
This work focussed on the topic of efficient rendering and interaction with terrain and city models. The challenges in this area are the immense size and complexity of the raw input data and to enable an intuitive, efficient user experience during interaction.
In this direction, I have examined the following aspects:
• scalable LOD methods for photorealistic realtime rendering resulting in a hierarchical structure of models tuned for rendering
• in the field of reconstruction: abstraction of raw input models, to reflect their inherent structure better.
• basic techniques to augment the purely geometric/radiometric visual models with arbitrary additional semantic data.
I will now shortly recall the main findings of the preceding chapters, mention their further development and indicate future directions.
7.1 Rendering of Digital Surface Models
The research presented in chapter2has proven that it is possible to render huge out-of-core terrain datasets in realtime with photorealistic detail. The quadtree hierarchy equipped with precomputed mesh approximations of known accuracy and corresponding texture images is the basis of the realtime rendering method.
The overheads of traversing the quadtree and auxilliary computations are so small that the costs of rendering a frame are dominated by the visible part of the model and thus scalability is guaranteed. At the same time, the strict approximation bounds based on the Hausdorff distance render the LOD switches almost unnotica-ble, which greatly enhances realism. The individual tiles of the quadtree can be
compressed very efficiently, which is a bonus in low-bandwidth scenarios such as internet based streaming.
While back in 2003, the largest available model consisted of about 256 million height samples and the maximal ground resolution was 2m, in the meantime I was able to process digital surface models with up to 10cm resolution and up to 20 billion height samples. Moreover, the rendering application based on the presented approach is still capable to render such datasets at realtime framerates.
The approach has been transferred to the enterprise 3D RealityMaps, where high resolution data of most of the Eastern Alps was acquired and preprocessed using my proposed method. Together with the corresponding viewer applica-tion, those regions are available as streaming terrain datasets athttp://www.
outdoor-guides.de/tourenplaner. The scalability of my method has even been established on thin clients, culminating in the release of the first public smartphone app with a high-detail 3d map of the Alpine Ski World Championships at Schladming in 2013.
Scientifically, there are three directions which seem fruitful:
Improved models. The most prominent defect in 2.5d modeling is the lack of texture information on steep parts. Although in terrain such features are very rare, they often represent very interesting parts or even famous landmarks. Schneider and Klein[2008] show how to semi-automatically add photos to the terrain surface.
They are able to texturize also steep parts by using a reparameterization of the corresponding tiles.
Another issue of purely geometric simplification is the lack of structure-awareness. I address this in chapters4and5, where I refine the LOD approach regarding the representation of models at coarser LODs in order to get semantically improved models.
Improved rendering. There are several ways to increase the photorealism of the rendered images. Rendering with textures is based on the assumption that the whole surface is lambertian diffuse, i.e. takes the same color from all perspectives.
Especially for cities, where the roofs dominate the orthogonal perspective but fa-cades become visible in slanted perspectives, this assumption can not be transferred to the simplified models.Ruiters[2008] presents a method to model coarser LODs with a compact BTF model instead of diffuse textures.
Further possible improvements in this direction are models for atmospheric scattering or rendering with high-dynamic range and tone-mapping.
Towards 3d GIS. The high accuracy of my method maintains georeferences exactly and is therefore well suited for GIS applications. Consequently, it has