yanx27/Pointnet_Pointnet2_pytorch

Question regarding Dataloaders for Semantic Segmentation using PointNet++

An-u-rag opened this issue · 1 comments

An-u-rag commented

Hi, I am slightly confused about a couple of things.

  1. Why are there two separate dataloaders for train_semseg and test_semseg in S3DISDataLoader.py? Both seem to use different techniques for subsampling of points.
  2. I am confused regarding the concept of hierarchical partitioning of input point clouds, there seems to be a sampling done in the dataloaders as well as the PointNetAbstraction layers. However, The PointNet++ paper only mentions furthest point sampling and a volumetric ball query around it. Why is that?
  3. The S3DISDataset seems to use a random sampling technique to get blocks of fixed number of points whereas ScannetDatasetWholeScene uses grid sampling to get a block of fixed number of points. Why is this done? Would this not give inaccurate results?
FSet89 commented

Any update on this?