@inproceedings {pub3636,
	title = {On the Potential and Challenges of Neural Style Transfer for Three-dimensional Shape Data },
	author = {Timo Friedrich AND Nikola Aulig AND Stefan Menzel},
	year = {2018},
	month = {September},
	abstract = {In the field of two-dimensional image and video processing, convolutional neural networks have been successfully applied to generate novel images by composing content and style of two different sources, a process called artistic or neural style transfer. However a usage of these methods for three-dimensional objects is not straight forward due to the unstructured mesh representations of typical shape data. Hence, efficient geometry representations are required to use neural network based style transfer concepts for three-dimensional shapes and to enable the fast creation of style options for instance in a product ideation process. Due to the availability of extensive databases containing three-dimensional volumetric models, like e.g. the Princeton{\textquoteright}s ModelNet, the advances in deep neural networks and increased computational power, a portfolio of computational methods has been proposed for shape classification and segmentation based on voxel, mesh, graph or point cloud representations. In this paper an overview of current state-of-the-art shape representations is presented with respect to their applicability of neural style transfer on three-dimensional shape data. Combinations of three-dimensional geometric representations with deep neural network architectures are evaluated towards their capability to store and reproduce content and style information based on previously proposed reconstruction tests. While mid- and high-resolution three-dimensional voxel models are computationally demanding towards required graphics memory, low voxel grid resolutions mask the shape{\textquoteright}s style characteristics. The utilization of processing point cloud data in neural networks usually incorporates strategies to become order invariant. However, this invariance which is highly beneficial during inference challenges the gradient back propagation during style transfer optimizations. Further experiments extend the analysis of the opportunities and allow a discussion of alternative approaches for realizing style transfer to three-dimensional objects and future applications.   },
	publisher = {Springer, LNCS, LNAI, LNBI},
	booktitle = {EngOpt 2018}
}