PaperReading

GAC: Ground-aware Monocular 3D Object Detection for Autonomous Driving

August 2021

tl;dr: Anchor-based method with a ground aware convolution module.

Overall impression

This paper is directly inspired by M3D-RPN. It still uses anchors instead of anchor-free, and uses the postprocessing module to optimize yaw.

Key ideas

• 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.

Technical details

• Summary of technical details

Notes

• Questions and notes on how to improve/revise the current work