The std of boundary sampling

[heathentw]

As in the training cmd, the project use 0.15 and 0.015 to multiply on a normal distribution for points shift augmentation; my question is that what size is your voxel (in meter) since those shift will directly add onto points sampled from mesh, and what parameter & voxel size will you use while training IPNetMANO? I assume that uniformly sample from voxel grid lead to a worse results and you guys tried some set of stds to get the best results.