HDNET/PIXOR++: Exploiting HD Maps for 3D Object Detection

November 2020

tl;dr: Exploiting HD map priors to guide 3D lidar object detection.

Overall impression

The paper is very well written with clear logic.

Maps contain geographic, geometric and semantic priors that re useful for many tasks. While HD maps are widely used by motion planning, they are vastly ignored by perception systems.

The PIXOR++ baseline can already beat previous lidar-only method such as PIXOR, MV3D and sensor fusion methods such as ContFuse.

Key ideas

Technical details