Fig 4. keypoints: from Farthest point sampling ⇒ Sectorized Proposal-Centric (SPC) keypoint sampling

uniformly sample keypoints from neighboring regions of proposals

分四个象限有啥好处？

Fig 5. VectorPool aggregation for local feature aggregation? why 有用吗？

inference speed: on a single TITAN RTX GPU

1. 10 FPS for 150m × 150m detection range on the Waymo Open Dataset (three times faster than PV-RCNN framework),
2. 16 FPS for 70m × 80m detection region on the KITTI dataset.
   1. PointPillars: 62 fps on KITTI on a desktop with an Intel i7 CPU and a 1080ti GPU, but 82.58 vs 90.14 for easy car 3d detection

PV-RCNN++: Point-Voxel Feature Set Abstraction With Local Vector Representation for 3D Object Detection