August 2021
tl;dr: Anchor-based method with a ground aware convolution module.
This paper is directly inspired by M3D-RPN. It still uses anchors instead of anchor-free, and uses the postprocessing module to optimize yaw.
- The key idea is a ground aware convolution (GAC) module. The network predicts the offers in the vertical direction and we sample the corresponding features and depth priors from the pixels below.
- Depth priors are inspired by CoordConv and are computed with perspective geometry with ground plane assumption.
- Summary of technical details
- Questions and notes on how to improve/revise the current work