Semantic3D lower result

[hpc100]

Thanks for the great work and open-sourcing it. @guochengqian
Has anyone tested PointNext on Semantic3d (reduced-8) ?
It seems the results are very weak (close to mIoU 50) and model are underfitting.

[guochengqian]

Have you played with radius and data augmentations.
Radius matter a lot for different datasets.
Please let me know

[hpc100]

@guochengqian thank's for the reply
Have you any ideas of the possible range of radius ?
With Semantic3D, I also have one other problem referenced in #138 ; concerning ignore_index in IoU introducing a gap between train and val IoU.
I will try to change radius.
Thank's.