Symmetry-based Shape Analysis and Modeling

Recent feature films and computer games (such as Avatar or Crysis) demonstrate impressively that computer graphics has reached a remarkable state-of-the-art: offline rendering has become practically indistinguishable from real-world photography and real-time rendering is getting closer as well. Nevertheless, the field is still facing significant challenges: Content creation has nowadays become the major bottleneck. Creating 3D models is still a tedious task that requires not just artistic talent but also significant technical skills. And even for highly skilled and trained professionals, modeling high quality 3D assets is extremely time consuming. This has led to a situation where 3D computer graphics is still used sparsely in practice, mostly being limited to large budget productions, such as high end game and movie projects.

Obviously, better tools are required to facilitate the task. Given that the field of 3D modeling is similarly mature as 3D rendering, we argue that a radically different approach is required to make a difference over the current state-of-the-art. We believe that one approach with the potential to drastically simplify 3D modeling is a data-driven approach to 3D modeling. We already have large 3D data bases of existing assets, as well as extensive data sets from 3D scanning (in this context, think for example of the Google street view project, which uses 3D LIDAR scanners for acquiring geometry along with images of all urban area on this planet).

We therefore need to analyze, structure, and, to some extent, understand the existing 3D geometry to reuse the available information for creating high quality content more rapidly: Parts, style, and compositions of existing content need to be made accessible in a suitable structure to be reused in the creation of novel content.

The goal of the proposed project is to develop new methods for finding structure in 3D shapes that can be used to assemble model variations from example geometry. Our approach consists of two parts: geometry analysis and synthesis, with the special case of compact storage and visualization.

  • Analysis: Our analysis approach is based on symmetry. This means, we analyze shapes for redundancy and detect parts that show up repeatedly within the same input. The analysis yields a set of building blocks that the input shapes are composed of, and from their layout, we can also infer rules how these building blocks can be used to assemble composite shapes.
  • Synthesis: In a second step, we use the discovered parts and rules to build new shapes. Here, it is particularly important to find expressive and controllable formal models that capture structure in example geometry. On the algorithmic side, we need efficient methods that are fast enough to support interactive editing and creation of shapes.
  • Compression and visualization: A strong generative structure model of 3D geometry will in particular be able to yield compact representations of the data. Therefore, we also investigate using our representations for compression, and efficient rendering of 3D models from compressed representations.

Project team

Principal Investigator
Dr. Michael Wand

Co-Principal Investigator
Prof. Dr.-Ing. Philipp Slusallek

Researchers
Javor Kalojanov
Chuan Li
Tomas Davidovic