Functionality Preserving Shape Style Transfer

摘要

When geometric models with a desired combination of style andrnfunctionality are not available, they currently need to be createdrnmanually. We facilitate algorithmic synthesis of 3D models of manmadernshapes which combines user-specified style, described via anrnexemplar shape, and functionality, encoded by a functionally differentrntarget shape. Our method automatically transfers the style of thernexemplar to the target, creating the desired combination. The mainrnchallenge in performing cross-functional style transfer is to implicitlyrnseparate an object’s style from its function: while stylisticallyrnthe output shapes should be as close as possible to the exemplar,rntheir original functionality and structure, as encoded by the target,rnshould be strictly preserved. Recent literature point to the presencernof similarly shaped, salient geometric elements as a main indicatorrnof stylistic similarity between 3D shapes. We therefore transfer thernexemplar style to the target via a sequence of element-level operations.rnWe allow only compatible operations, ones that do not affectrnthe target functionality. To this end, we introduce a cross-structuralrnelement compatibility metric that estimates the impact of each operationrnon the edited shape. Our metric is based on the global contextrnand coarse geometry of evaluated elements, and is trained onrndatabases of 3D objects. We use this metric to cast style transfer asrna tabu search, which incrementally updates the target shape usingrncompatible operations, progressively increasing its style similarityrnto the exemplar while strictly maintaining its functionality at eachrnstep. We evaluate our framework across a range of man-made objectsrnincluding furniture, light fixtures, and tableware, and performrna number of user studies confirming that it produces convincingrnoutputs combining the desired style and function.