Frame semantics in text-to-scene generation

Abstract

3D graphics scenes are difficult to create, requiring users to learn and utilize a series of complex menus, dialog boxes, and often tedious direct manipulation techniques. By giving up some amount of control afforded by such interfaces we have found that users can use natural language to quickly and easily create a wide variety of 3D scenes. Natural language offers an interface that is intuitive and immediately accessible by anyone, without requiring any special skill or training. The WordsEye system (http://www.wordseye.com) has been used by several thousand users on the web to create over 10,000 scenes. The system relies on a large database of 3D models and poses to depict entities and actions. We describe how the current version of the system incorporates the type of lexical and real-world knowledge needed to depict scenes from language. © Springer-Verlag 2010.