3D Scene Analysis

Thorsten Thormählen

3D Scene Analysis

Simple creation of new 3D objects through blending of existing model

On the Internet, more and more free or commercial databases of professionally- created 3D objects are becoming available. For a non-expert user, the complex operation of current modeling tools makes it difficult to create similarly good 3D objects. Therefore, often the only option these users have is to choose from a 3D database an existing model that most closely matches their vision. As part of our project, a tool was developed that makes it possible for a non-expert user to create a new 3D object according to his or her wishes by combining existing objects from a 3D database. To that end, the database objects must be automatically analyzed to determine which parts can be exchanged. The analysis includes firstly the segmentation of an object into its component parts. Next, contacts and symmetries among the parts are identified. The process then enables the creation of variants of the object that have similar contacts and symmetries.

Figure 1:
Automatic creation of a blend (center) between a bicycle (left) and a motorcycle (right).

Automatic assignment of materials for 3D objects

The material composition of a 3D object is strongly correlated with the object’s geometric shape and its relation to other objects in its immediate vicinity (context). In a further project, we have sought to model these context-dependent correlations and to show that they can be learned from a database containing several hundred 3D objects with known materials. If a 3D model without materials is then passed to the software we have developed, the learned model can be used to assign plausible material parameters fully automatically (including diffuse and specular reflection, gloss, and transparency).

Figure 2:
Given a 3D object without materials (left), our approach can automatically assign appropriate
materials (second from left). It also suggests alternative materials to the user; these can be selected interactively
to improve the automatic assignment (right).

We have also developed a user interface that provides suggestions for alternative materials. This user interface can be used to further refine the automatic suggestions [see figure 2 right]. Once a refinement has been made, the model uses this information to improve the automatic assignment of other parts of the 3D object. The approach works for 3D objects with different numbers of parts and with different topological complexity. Studies with non-expert users showed that the suggested process simplifies and accelerates the assignment of materials significantly in comparison to conventional approaches.