BEV-IPM: Deep Learning based Vehicle Position and Orientation Estimation via Inverse Perspective Mapping Image

October 2019

tl;dr: IPM of the pitch/role corrected camera image, and then perform 2DOD on the IPM image.

Overall impression

The paper performs 2DOD on IPM’ed image. This seems quite hard but obviously doable. The GT on BEV image seems to come from 3D GT, but the paper did not go to details about it.

The detection distance is only up to ~50 meters. Beyond 50 m, it is hard to reliably detect distance and position. –> Maybe vehicle yaw are not important for cars beyond 50 meters after all?

Key ideas

Technical details