Global-to-Local Generative Model for 3D Shapes

摘要

We introduce a generative model for 3D man-made shapes. The presentedmethod takes a global-to-local (G2L) approach. An adversarial network(GAN) is built first to construct the overall structure of the shape, segmentedand labeled into parts. A novel conditional auto-encoder (AE) is then augmentedto act as a part-level refiner. The GAN, associated with additionallocal discriminators and quality losses, synthesizes a voxel-based model, andassigns the voxels with part labels that are represented in separate channels.The AE is trained to amend the initial synthesis of the parts, yielding moreplausible part geometries. We also introduce new means to measure andevaluate the performance of an adversarial generative model. We demonstratethat our global-to-local generative model produces significantly betterresults than a plain three-dimensional GAN, in terms of both their shapevariety and the distribution with respect to the training data.