Geometric modeling is essential for populating virtual environments and designing real objects, ranging from furniture and car assembly to 3D modeling of chemical compounds and medical devices. Creating 3D models from scratch is a tedious and time-consuming process that typically requires expertise from the person building them.
Researchers from Stanford University, University of California at San Diego, Adobe Research and IIT Bombay have collaborated on a novel computational framework for assembly-based 3D modeling that automatically suggests to users which parts to use and where to place the part in the actual design.
Given a partial object, for example, a basic chair back, the team’s method proposes a set of potential complementary components from a repository of 3D shapes, like the seat and chair legs. In examples, the researchers were able to show that their method automatically suggests complementary parts based on any shape proposed. The software suggests placement in the assembly as well. For instance, in designing a desk, the method can automatically predict whether the next step in the process is to add a drawer, and furthermore, whether that drawer should be placed in the center of the table, the left-hand side or the right-hand side.
If a user wants more design control, the method provides an interactive capability. Given the partial assembly of part of a sofa, or a basic shape of a sofa, the algorithm first proposes a set of possible components. Then the user gets to choose which component to use in their design. The method predicts where the piece is placed. When evaluated, the new approach demonstrated significant improvement over state-of-the-art retrieval techniques.
"Our method leads to a modeling tool that requires minimal or no user inputs," said Minhyuk Sung, lead author and a computer science Ph.D. student at Stanford, advised by coauthor Leonidas Guibas, professor of computer science and electrical engineering at Stanford. "We've come up with novel solutions to two key technical components in assembly-based modeling: part retrieval and part placement."
A key contribution of the research is that the method allows users to pull unlabeled data obtained from anywhere on the internet, making the process of automating 3D model design more efficient. Existing techniques for computational design of 3D models typically include a detailed step for labeling and indexing parts required for the design. The new technique doesn’t require the models used to train the algorithm to be consistently segmented into labeled parts. Instead, it can learn from un-annotated inconsistently segmented collections of models in online repositories.
In the future, researchers plan to augment their framework with capabilities to further manipulate and customize retrieved parts, providing even better compatibility with the initial query. To make the modeling process even more interactive and intuitive, the researchers intend to add more fine-grained suggestions.
The paper on this research was presented by the authors at SIGGRAPH Asia 2017 in Bangkok in November.