GUPNet: Geometry Uncertainty Projection Network for Monocular 3D Object Detection

August 2021

tl;dr: Uncertainty prediction of 3D height transfer to uncertainty of depth.

Overall impression

The multi-task learning part is quite interesting, but the depth prediction part lacks clarity and insight (it is more like a post-hoc experiment report).

The relationship between height and depth indeed can be mined, but with a sutble difference that the projected $h_{3d}$ does not necessarily match the height of the 2D bbox $h_{2d}$.

Key ideas

Technical details