Accurate Monocular 3D Object Detection via Color-Embedded 3D Reconstruction for Autonomous Driving

August 2019

tl;dr: Concurrent proj with Pseudo-lidar but with color embedding. SOTA together with CMU PL for mono 3DOD as of 08/2019.

Overall impression

This paper seems to be inspired by pseudo-lidar but did not mention it explicitly. The paper’s writing leaves much to be desired. Performance-wise, it is very similar to CMU’s e2e pseudo-lidar paper.

Simply concat D to RGB does not help much, such as ROI 10D and MLF.

The paper proposed a simple segmentation method to combat the long-tail problem, although it does not clearly state so.

Key ideas

Technical details