Learning-Deep-Learning

BS3D: Beyond Bounding Boxes: Using Bounding Shapes for Real-Time 3D Vehicle Detection from Monocular RGB Images

August 2020

tl;dr: Annotate and predict a 6DoF bounding shape for 3D perception.

Overall impression

The paper proposed a way to annotate and regress a 3D bbox, in the form of a 8 DoF polyline.

This is one of the series of papers from Daimler.

• MergeBox
• Bounding Shapes
• 3D Geometrically constraint keypoints

Key ideas

• Bounding shape is one 4-point and 8DoF polyline.

Technical details

• The paper normalizes dimension and 3D location to that y = 0. When real depth is recovered (via lidar, radar or stereo), the monocular perception
• An object is of Manhattan properties if 3 orthogonal axes can be inferred, such as cars, buses, motorbikes, trains, etc.

Notes

• 57% of all American drivers do not use turn signals when changing the lane. (source)