Issues about input region points and augmentation part.

[JudgeLJX]

Hi Hugues,

Thanks for your excellent contribution.

I have some questions.

When I want to handle the S3DIS dataset, the first thing is subsampling it to yield a subsampled point cloud for example area4_sub.ply, I use the room as input here. And for the next step, it will choose a center point with a fixed radius to generate the input sphere region. For this sphere region,

Q1: Are all points in these regions normalized to a fixed scale, only coordinate information normalized, or both xyz and rgb (though only y as input)?

Q2: How to guarantee that we can traverse all points in the point cloud? I found you apply a potential probability to choose the center point, but it seems still some points will not be involved.

And for another part about data augmentation in the dataset, there is an augmentation transform to the input point which changes the xyz coordinate,

Q3: If there is no augmentation step, the local information of any center point may vary, will this operation lead to the change of samples, pools, and neighbors in input_list?

Best Regards,

Judge