Regarding the scale of the ground truth point maps in training.

[GonzaloMartinGarcia]

Hi, congrats on the amazing results!

I have a question regarding the scale of the ground truth point map when it is aligned with the prediction in the ROE solver. From the paper, I assume that the ground truth is provided in metric scale, and that using inverse depth (1/z) for the weights helps unify differences between indoor and outdoor scenes. Are there any additional normalization steps applied to the ground truth before the alignment, such as centering or bounding box normalization, or is the alignment done directly in metric scale?

This would also influence how one chooses the \tau hyperparameter parameter, given that scene scale differences might affect when their errors are clipped.

Thanks and regards.