Learning-Deep-Learning

Mono3D++: Monocular 3D Vehicle Detection with Two-Scale 3D Hypotheses and Task Priors

August 2019

tl;dr: Mono 3DOD based on 3D and 2D consistency, in particular landmark and shape recon.

Overall impression

The paper is written in overcomplicated math formulation. Overall not very impressive. The consistency part is quite similar to other papers such as deep3dbox.

The morphable wire frame model is fragile and the authors did not do a thorough ablation study on its contribution. I am not sure if shape recon is a good idea, especially to handle corner cases. –> Nobody in the literature actually talks about how to handle corner cases. This need to be acquired through engineering practice. Maybe CV method is needed to handle the corner cases.

The paper seems to use 3D depth off the shelf but it was not described in details.

Key ideas

Technical details

Notes