The aim of the MoSeS project is to use Artificial Intelligence techniques to infer semantic properties of shapes

I3D visualization and modeling have proved to be effective in the design of Engineering applications, product realizations, pilot training, ergonomic evaluation, simulation of complex system behavior and even surgery. Modern Industrial Design can no longer be imagined without the support of Computer Aided Design (CAD) tools and digital 3D models. Therefore semantic processing and interpretation of shapes and surfaces in CAD systems is an important problem.

The MoSeS project is primarily concerned with attaching semantics (textual, hierarchical and geometrical) automatically to shapes so the annotated models can be archived, queried and shared. To achieve this, we propose a generic architecture which includes feature extraction, classification and knowledge management modules.

Feature Extraction: Recently, numerous systems have been developed to search the Web for 3D models which resemble some query object (sketched or uploaded as a full geometry). To do this, generic feature extraction methods were developed. These are invariant under rotation, translation and scaling to account for the versatility of models that are posted on the Web. For our purposes, however, these methods are too general and therefore we are investigating new ways of feature extraction techniques which can incorporate domain knowledge.

Classification: Currently the well known "nearest neighbor" algorithm is the most popular to establish shape similarity based entirely on features extracted from geometry. Annotated shapes, however, will contain discrete (e.g. textual) tags as well as quantitative features only partially derived from geometry and these will provide the opportunity to use more sophisticated classifiers, such as, Decision Trees, Neural Networks and Support Vector Machines.

Knowledge Management: To be able to use, archive and retrieve shapes according to semantic features, ontologies of the domain (products and assemblies) are needed. For MoSeS, we are also building a "shape ontology" that will be used to describe objects with respect to physical, mathematical and aesthetic properties.