Learning-Deep-Learning

LiDAR R-CNN: An Efficient and Universal 3D Object Detector

September 2021

tl;dr: Add proposal info to the point cloud before feeding into the 2nd stage.

Overall impression

PointNet could make the learned features ignore the size of proposals, as it only aggregates the features from points while ignoring the spacing in 3D space. The spacing encodes essential information such as the scale of the objects.

Lidar RCNN provides a plug-and-play module to any existing 3D detector to boost performance. –> This could be useful for offline perception.

It is a point-based method, like Point RCNN and PV RCNN.

Key ideas

Technical details

Notes

-code on github