efficiency

[lixinghe1999]

Dear author,

From my understanding, the fusion of 3D bbox and 2D bbox is very inefficient, even you only process the non-empty entries on the deep learning model.
Specially, I find the 3D -> 2D conversion, 2D - 2D IoU calculation are very slow, even though @numba is used.
Do you have any hints or I misunderstand something?
Actually I don't use this repo but copy the code from it because I want to use spconv 2.0, which is too heavy to transfer this repo to newest version. As a result, I implement CLOCS on OpenPCDet by myself. I am not sure my conclusion is 100% correct.