mb_net.md

MB-Net: MergeBoxes for Real-Time 3D Vehicles Detection

August 2020

tl;dr: Use 5DoF 2d bbox to infer 3d bbox.

Overall impression

The paper proposed a way to annotate and regress a 3D bbox, in the form of a 5 DoF bbox (MergeBox).

This is one of the series of papers from Daimler.

• MergeBox
• Bounding Shapes
• 3D Geometrically constraint keypoints

Key ideas

• A 5 DoF bbox to represent 3d bbox.
• 3D car size templates have to be assumed to lift the mergebox representation to 3D.

Technical details

• The authors noted that even one single template can achieve good performance for AOS (average orientation score).

Notes

• The fancy name for (cos(theta), sin(theta)) is called Biternion. The gaussian on unit circle is called von Mises distribution.
• 3D annotation generally has two approaches: using lidar or 3D CAD model.
• This is similar to what nvidia does by marking the visible edges of the car.