没谈motion的影响，why？
加上估计的motion会更好吧。
semi-supervised learning (SSL) in LiDAR segmentation
leverage the strong spatial cues of LiDAR point clouds to better exploit unlabeled data.
we achieve competitive results over fully-supervised counterparts with 2× to 5× fewer labels and improve the supervised-only baseline significantly by 10.8% on average. 关键是比Cylinder3D好了多少？没说！！
LiDAR scans contain strong spatial prior. Objects and backgrounds around the egovehicle have a patterned distribution on different (lower, middle, upper) laser beams.
mixes the aforementioned laser partitioned areas A from two scans in an intertwining way, i.e., one takes from odd-indexed areas A1 = {a1, a3, ...} and the other takes from even-indexed areas A2 = {a2, a4, ...}, so that each area’s neighbor will be from the other scan:
split point cloud into 8 non-overlapping areas, i.e., A = {a1, a2, ..., a8}. Each ai contains points captured from the consecutive 8 laser beams.
mixing-based teacher-student training pipeline, 其实很普通
only use the Teacher net during inference