todo: 理解一下它的former。

a little higher performance than RangeFormer on nuScenes and semanticKitti. 和lidar表示相关的transformer，这比RangeFormer好。

Varying-sparsity property of LiDAR point clouds. The dense close car is marked with a green circle and the sparse distant bicycle is marked with a red circle

1. Dynamic feature selection. We split the heads to conduct radial and cubic window self-attention respectively.

Reference

1. SphereFormer: Spherical Transformer for LiDAR-based 3D Recognition, cvpr23