Centroid Voting: Object-Aware Centroid Voting for Monocular 3D Object Detection

December 2020

tl;dr: Use bbox to guide depth prediction.

Overall impression

The paper is really a run-of-the-mill paper. The main idea is that instead of convolutional features to regress distance, use geometric prior to guide the distance prediction. The convolutional appearance features are only required to learn the residual.

Key ideas

Technical details