@inproceedings {pub4444,
	title = {The effects of non-linearity operators in voxel-based 3D neural networks for shape reconstruction},
	author = {Timo Friedrich AND Patricia Wollstadt AND Stefan Menzel},
	year = {2020},
	month = {November},
	abstract = {Neural Style Transfer has been successfully applied for generating a plausible novel 2D image based on existing image content and the style of a painting, with further extensions to image, video, motion and animation manipulation. Follow up works have realized 2D/3D hybrid approaches relying on 2D rendering operations for first steps towards 3D shape manipulation. However, for an end-to-end 3D Neural Style Transfer we suggested 3D voxel-based neural network architectures, which we extend in the present paper to 256^3 voxel resolution to depict, encode and reconstruct vastly finer visual features. In a series of experiments, we have identified that the type of activation functions used in the neural network classification models is crucial for the successful application of these networks in later Neural Style Transfer use cases. Hence, we have trained and compared multiple network setups with different non linearities including the novel bipolar exponential linear unit (belu) operator. Based on our standardized gram matrix approach and further previously found network optimizations, we then reconstruct purely style-based shapes. We demonstrate the positive impact of the belu activation functions compared to the commonly used exponential linear units (elu) and provide insights for the shown behavior. The constructed shapes are eventually evaluated using local shape descriptors which confirm the subjective visual improvements. We conclude that the introduction of bipolar exponential linear units into the style transfer process is a crucial component on the way to realize high resolution voxel-based style transfer.},
	publisher = {IEEE},
	booktitle = {IEEE Symposium Series on Computational Intelligence (SSCI)}
}