search

JonasSchult / Mask3D

Mask3D predicts accurate 3D semantic instances achieving state-of-the-art on ScanNet, ScanNet200, S3DIS and STPLS3D.
MIT License
547 stars 110 forks source link

about s3dis conf #53

Closed liujingangel closed 1 year ago

liujingangel commented 1 year ago

Hello, thans for your work! when I learn the s3dis config to train model on my custom data, I'm confused why "data.num_labels(13)" is less than "general.num_targets(14)", and other datasets equal. https://github.com/JonasSchult/Mask3D/blob/b60b143ceca2e372bac9147d112af1dc6dccb7d3/scripts/s3dis/s3dis_from_scratch.sh#L14

can you please help me with this confusion? thanks in advance

JonasSchult commented 1 year ago

Hi!

thanks for your interest in our work!

In S3DIS, every point is associated with an instance (e.g. wall, floor are instances in S3DIS). This is in contrast to ScanNet, where wall, and floor are not treated as instances.

In particular for S3DIS, we have 13 classes but use 14 targets (+1 for the no-instance semantic class).

Best, Jonas