Question: Is the RGB in the .las files real sensor color or semantic pseudo-color?
Hi, thank you for releasing TowerDataset! Before using it, I'd like to clarify one thing.
Each .las file (LAS 1.2, point format 2) contains XYZ coordinates along with RGB, while semantic labels are stored in the classification field.
Could you confirm whether this RGB is:
Real sensor-acquired color , or
Pseudo-color rendered from the semantic class (each point colored by a fixed RGB mapped from its classification label)?
Thanks!
Hi, thanks for checking! The RGB fields should not be treated as a consistent sensor-acquired color modality. Some files contain semantic pseudo-colors assigned from class labels for visualization.
Please use the classification field as the ground-truth semantic label. Our TowerDataset experiments use XYZ coordinates only and exclude RGB. We recommend ignoring RGB when training or evaluating segmentation models, as class-derived colors can leak the ground-truth labels