monoDLE: Delving into Localization Errors for Monocular 3D Object Detection

April 2021

tl;dr: Boost the performance of accurate 3D localization by using 2D bbox prediction task as auxiliary task, and discarding bad samples.

Overall impression

This paper has a very thorough analysis into the error of mono3D. This reminds me of the What monodepth see ICCV 2019 paper. It founds that localization error is one vital factor accounting for the poor performance of mono3D.

In addition, accurately localizing distant objects with existing technologies is almost impossible. So removing those distant objects can boost performance.

The 3D location of an object can be decoupled into two parts, projected center of the 3D objects, and the depth.

Key ideas

Technical details