HDMapNet: An Online HD Map Construction and Evaluation Framework

July 2021

tl;dr: Structured static BEVNet with optional lidar camera early fusion.

Overall impression

The paper focuses on prediction of vectorized map elements in the birds-eye-view for scene understanding. The perspective to BEV transformation is not done via IPM but rather with a MLP. The extrinsics are only used to piece the cameras together to the ego frame (thus with only translation, not rotation).

The static BEVNet is essentially an online map. learning framework that makes use of onboard sensors and computation to estimate local maps.

Key ideas

Technical details