Learning-Deep-Learning

PointRCNN: 3D Object Proposal Generation and Detection from Point Cloud

Mar 2019

tl;dr: Extend existing point cloud classification and segmentation method to 3D detection (instance segmentation and amodal bbox estimation). The PointRCNN directly generates high quality proposals from point cloud.

Overall impression

This paper basically extended Faster RCNN and Mask RCNN to point cloud representations, with many tweaks to adapt the representation difference. This work is closely related to other 3D detection framework such as Frustum PointNet, in particular the study in the appendix where proposals are generated from bird’s eye view from point cloud. The main contribution is the first stage which generates high quality 3D bbox proposals from point cloud.

The idea of predicting one bbox per point is also used in LaserNet.

Key ideas

Technical details

Notes