Learning-Deep-Learning

MonoFlex: Objects are Different: Flexible Monocular 3D Object Detection

April 2021

tl;dr: Decouple the prediction of truncated objects in mono3D.

Overall impression

How to deal with truncated objects remained one key task for mono3D. Most of the existing method (especially CenterNet) did not treat this in particular and the prediction on truncated objects looks distorted. This is the first paper which explicitly addresses this challenge.

The idea is to decouple the learning of truncated objects (outside objects) and untruncated object (inside objects), by using different representative points (or anchor point).

Key ideas

Technical details

Notes