Question about S3DIS Area_5b 2D-3D projection / pose alignment

[MengHaobo]

Hi, thank you for releasing the code and checkpoints.

I am trying to reproduce the S3DIS results following the README. I used the original Stanford3dDataset_v1.2 point clouds and the 2D-3D-S no_xyz data for RGB/depth/pose.

The official S3DIS checkpoint gives results very close to the paper on my local setup:

• Paper: mIoU 46.5
• Official checkpoint evaluated locally: mIoU 46.41

However, when I regenerated the DINOv2 projected features myself using the provided projection code, my reproduced training result was much lower. While checking the projected PCA .ply files and 2D-3D visibility, I noticed that some Area_5b panoramas seem difficult to align with the
Area_5 point clouds. In a few cases, the projection produced very few or almost zero visible 3D points.

I am not sure whether this is a dataset/preprocessing issue on my side or whether Area_5b needs any special handling. I noticed that another S3DIS-related project, DeepViewAgg, applies a manual calibration for Area_5b poses. I was wondering whether you encountered any similar Area_5b
pose/alignment issue when preparing the projected DINOv2 features for LogoSP.

Could you please clarify:

1. Did you apply any special pose correction or manual calibration for Area_5b when generating the released S3DIS projected DINOv2 features?
2. Did you use both Area_5a and Area_5b in the distillation stage for the released S3DIS checkpoint?
3. From the code, it looks like S3DISdistill uses all areas by default, while train_Seg_S3DIS.py excludes Area_5 only during the segmentation training stage. Is this the intended setting for reproducing the paper?

Thanks in advance!