why does the 3D Position Encoder conv 3x3 not work?

[Alexanderisgod]

Your paper says that the conv 3x3 destroy the correspondence between 2D image feature and 3D position. However, from the point cloud view, image lacks the ability to discovery 3D structure behind the 2D plane. If step further, could it be possible to transfer the point cloud knowledge to image feature?