This codebase almost implements(unofficial) the system described in the paper:
PackNet-SfM: 3D Packing for Self-Supervised Monocular Depth Estimation
V. Guizilini, R. Ambrus, S. Pillai, and A. Gaidon
[ pdf, video, ICML'19 SSL wkshp ]
As there are many differences compared to original paper and the author released their training and testing code, I deleted code at master branch but keep going at dev branch, you can go to packnet-sfm(code published) for all detail.