CasGeo: 3D Bounding Box Estimation for Autonomous Vehicles by Cascaded Geometric Constraints and Depurated 2D Detections Using 3D Results

October 2019

tl;dr: Extends Deep3DBox by regressing the 3d bbox center on bottom edge and viewpoint classification.

Overall impression

CasGeo predicts more geometric keypoints or properties from the image patch. It regresses projection of the center of 3D bbox bottom face (CBF) and viewpoint classification.

It uses similar constraint such as GS3D to regress the 3D height of the bbox in 2D image. But instead of using the statistical mean of 93% of 2d bbox height, CasGeo regresses this distance from image patch. Then there are two steps: first approximate the location such as GS3D does, then feed into the Gauss-Newman method to solve the over-constraint problem proposed by deep3dbox.

The paper also removes false positive in 2D detection based on inferred 3D information.

Key ideas

Technical details