Monocular 3D Object Detection: An Extrinsic Parameter Free Approach

August 2021

tl;dr: Regresses the extrinsics and uses feature transfer to compensate.

Overall impression

This approach correctly addresses one drawbacks from existing mono3D dataset which assumes a fixed extrinsics. This is not true in industry applications due to potholed and uneven roads. The paper introduces a method to relax the constraint of assuming a fixed extrinsics.

Personally I am not super confident about the approach to regress extrinsics with horizon and vanishing point. This may work on highways but in crowded urban scenario this may fail miserably.

Key ideas

Technical details