3D Shape Synthesis via Content–Style Revealing Priors

Abstract To increase the diversity of created shapes conveniently, we describe a fully automatic synthesis framework based on the content–style analysis of an input set of 3D shapes. To fully characterize the content and style of the input set and enable 3D shape synthesis, we need a set of elements that capture a more complete characterization of the content and the style, which is precisely defined as the set of content–style revealing priors. First, the 3D shape priors extracted from the input collection of shapes are used to compute the global shape descriptors of each shape, and then cluster them to form the different content classes of the input shapes. We pose our search as a content–style analysis problem, where the separation between content and style would lead us to the revealing priors we seek. To this end, we introduce the prior-shape matrix (PSM) which measures how far a specific prior (row) appears in a specific shape (column). This allows us to classify the 3D sparse priors into content and style revealing priors. The results of the content–style separation phase are used to automatically synthesize novel shapes by transferring the style parts of the input shapes to their content parts while maintaining their functionality. Extensive experiments on several datasets of man-made 3D shapes show the identification of the content–style revealing priors and exhibit the power of our approach in synthesizing plausible shapes based on the separated styles and contents. In addition, our method enables a variety of applications such as content–style aware shape retrieval and stylistic suggestions for scene modeling, concluding that the revealing priors are a concrete description of how to grant a specific content or a style to an object.