Abstract Description Refinement Using Incremental Learning and Scene Reconstruction

Declarative Modeling methodologies offer the designer the ability to describe a scene using abstract terms instead of precise geometric elements and properties. The price for this convenience is a large number of compliant geometric models, only a small subset of which is usually of practical interest for the designer. The task of solution evaluation can be tedious and time-consuming whereas the qualities that make these solutions stand out are not always straightforward. In the current work we outline the integration of a machine learning component, trained by user-approved solutions of previous descriptions, with a reconstruction component, able to discover relations and properties implied by the best solutions, into a unique module for description adaptation according to user preferences.