3D Modeling in a Web Browser to Formulate Content-Based 3D Queries Ren´ e Berndt 1 Sven Havemann 1 Dieter Fellner 1,2 1 Institute of ComputerGraphics and KnowledgeVisualization (CGV), TU Graz, Austria 2 GRIS, TU Darmstadt & Fraunhofer IGD, Darmstadt, Germany Abstract We present a framework for formulating domain-dependent 3D search queries suitable for content-based 3D search over the web. Users are typically not willing to spend much time to create a 3D query object. They expect to quickly see a result set in which they can navigate by further differentiating the query object. Our sys- tem innovates by using a streamlined parametric 3D modeling en- gine on both client and server side. Parametric tools have greater expressiveness, they allow shape manipulation through a few high- level parameters, as well as incremental assembly of query objects. Short command strings are sent from client to server to keep the query objects on both sides in sync. This reduces turnaround times and allows asynchronous updates of live result sets. CR Categories: I.3.2 [Computer Graphics]: Graphics Systems— Distributed/network graphics; I.3.5 [Computer Graphics]: Comput. Geometry & Object Modeling—Geometric algor., languages, syst. Keywords: 3D search interfaces, rapid shape design, domain- dependent modeling tools, generative modeling, GML 1 Introduction The great success of content-based search and retrieval has become obvious with textual documents: Google & Co. allow looking for texts simply by specifying words that appear in these texts. This is a great paradigm change compared to the long-standing prac- tice in libraries: Traditionally a librarian ingests a book who un- derstands the text and assigns appropriate subjects and keywords typically chosen from a controlled standard vocabulary. In every library the subject index permits retrieving relevant text documents efficiently if, and only if, also the researcher is acquainted with the controlled vocabulary, or can use a thesaurus containing keyword synonyms. So neither assigning keywords nor retrieving texts is trivial. Content-based retrieval is easier for casual users, and it is also very transparent: Only documents that contain the query string are returned. The difficulty is the ranking, e.g., when thousands of documents match a query. Sophisticated ranking algorithms are used, e.g., to infer the intended context from the word combination in a multi-word query. Relevance feedback improves the precision by logging the results that are selected most often. Such levels of sophistication can only be achieved when content- based search and retrieval are available in the first place. But this is difficult for multimedia documents: Images, videos, music, drawings, and 3D models can be understood as generalized docu- ments. Very large digital libraries of such multimedia documents are manageable only if services for markup, indexing, and retrieval [Endres and Fellner 2000] are available. However, to formulate a content-based multimedia query users have to create an image, a video, or some music etc. For 3D-documents, the user must pro- vide a 3D model to formulate, e.g., a query-by-example. It turns out that “formulating” such a query can be a difficult problem. 1.1 Formulating a 3D-query Query-by-example requires a query object: When looking for chairs, the user has to provide a 3D model of a chair. Unfortu- nately, users typically look for models that they do not have. In this case, a new query model needs to be created, which leads to the difficult 3D modeling problem: High-end modeling toolkits such as Blender, AutoCAD, 3D-Studio Max, Maya etc. are out of scope for average or casual users. 3D modeling tools for non-expert users, e.g., Google SketchUp [Google 2008], provide a small but well- defined set of modeling tools and are easier to learn. But even with those tools creating a query model takes significantly longer than, e.g., typing in a text-based query. 1.2 Domain-dependent Parametric 3D Query Objects 3D search technology is generally better in discerning classes of objects (furniture, cars, animals) than in discriminating the objects within a class (Regency vs. Art Deco style, dog vs. cat, car model), as pointed out by [Bustos et al. 2007]. Shape subtleties are more difficult to express and take longer to sketch, so designing a query object requires more advanced skills. This problem can be solved using class- or domain-dependent parametric modeling tools: • Complex query objects are easier to create and to control us- ing only a small number of high-level parameters, • Subtleties can be expressed if there are specific parameters or modeling tools for them, • Search is more specific, e.g., for the number of floors of a building, the slope of a stairway.