Semantic segmentation input pointcloud size 8192

[dong-hub-png]

hi!
Thank you for your contribution. convpoint is a excellent work.I would like to ask if I need to do downsampling work if I use my own dataset, which can reach a single frame data volume of five million points? Or do I need to adjust the value of 8192 for point cloud input?

[aboulch]

I think best would be to:

1. voxelize to remove high density area
2. if the scene it large, cut it into pillars
3. decimate to 8192 into each pillar
   But there may be other ways to do it.

[dong-hub-png]

Thank you for your suggestion！I will verify these methods.