https://arxiv.org/pdf/2203.05662.pdf
faster & better than farthest point sampling for sparse points
use KDE to estimate local feature density, 是一种高级的说法。
Waymo Open Dataset [30], the detection range is [-75.2m, 75.2m] for the X and Y axes, and [-2m, 4m] for the Z axis. We divide the raw point cloud into voxels of size (0.1m, 0.1m, 0.15m). ⇒ 1500150040
Limitations
At higher voxel resolutions, the difference between the voxel center and the voxel point centroids decreases, resulting in less performance gains.
Point density-aware voxels for LiDAR 3D Object Detection, cvpr22, has code, with hash table. density算的快不？
Cirrus: A long-range bipattern lidar dataset. ICRA, 2021.
3d object detection with pointformer. CVPR, 2021.