The paper introduces PointRCNN, a novel framework for 3D object detection from raw point cloud data. The framework is composed of two stages: the first stage generates 3D proposals from the point cloud in a bottom-up manner, while the second stage refines these proposals in canonical coordinates to obtain final detection results. Unlike previous methods that project point clouds to bird's view or voxels, PointRCNN directly processes the point cloud to generate high-quality 3D proposals. The second stage transforms the pooled points of each proposal to canonical coordinates, combining them with global semantic features to refine the proposals and predict box confidence. Extensive experiments on the KITTI dataset show that PointRCNN outperforms state-of-the-art methods with significant margins, achieving the best performance among all published works as of November 16, 2018. The code for PointRCNN is available at https://github.com/sshao shuai/PointRCNN.The paper introduces PointRCNN, a novel framework for 3D object detection from raw point cloud data. The framework is composed of two stages: the first stage generates 3D proposals from the point cloud in a bottom-up manner, while the second stage refines these proposals in canonical coordinates to obtain final detection results. Unlike previous methods that project point clouds to bird's view or voxels, PointRCNN directly processes the point cloud to generate high-quality 3D proposals. The second stage transforms the pooled points of each proposal to canonical coordinates, combining them with global semantic features to refine the proposals and predict box confidence. Extensive experiments on the KITTI dataset show that PointRCNN outperforms state-of-the-art methods with significant margins, achieving the best performance among all published works as of November 16, 2018. The code for PointRCNN is available at https://github.com/sshao shuai/PointRCNN.