Interactive Design of Probability Density Functions for Shape Grammars

摘要

A shape grammar defines a procedural shape space containing arnvariety of models of the same class, e.g. buildings, trees, furniture,rnairplanes, bikes, etc. We present a framework that enables a user torninteractively design a probability density function (pdf) over suchrna shape space and to sample models according to the designed pdf.rnFirst, we propose a user interface that enables a user to quicklyrnprovide preference scores for selected shapes and suggest samplingrnstrategies to decide which models to present to the user to evaluate.rnSecond, we propose a novel kernel function to encode the similarityrnbetween two procedural models. Third, we propose a frameworkrnto interpolate user preference scores by combining multiple techniques:rnfunction factorization, Gaussian process regression, autorelevancerndetection, and l_1 regularization. Fourth, we modify thernoriginal grammars to generate models with a pdf proportional tornthe user preference scores. Finally, we provide evaluations of ourrnuser interface and framework parameters and a comparison to otherrnexploratory modeling techniques using modeling tasks in five examplernshape spaces: furniture, low-rise buildings, skyscrapers, airplanes,rnand vegetation.