Deep MANTA: A Coarse-to-fine Many-Task Network for joint 2D and 3D vehicle analysis from monocular image

CVPR 2017. 

https://arxiv.org/abs/1703.07570