Sparse Generative Neural Networks

@inproceedings{dai2020sg,
  title={Sg-nn: Sparse generative neural networks for self-supervised scene
         completion of rgb-d scans},
  author={Dai, Angela and Diller, Christian and Nie{\ss}ner, Matthias},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and
             Pattern Recognition},
  pages={849--858},
  year={2020}
}

Scene completion: partial TSDF -> reconstructed TSDF

use sparse convolutions

The later upsampling layers are only added one by one while training, such that we can make sure that the lower-res predictors are correct and that the higher res predictors are getting the correct input