Confidence loss and valid pixels for sky regions

Hello, thanks for the amazing work! I have a question about the confidence loss introduced in the paper and code. You mentioned that you compute it only over valid pixels. If I understated correctly, for example, the sky points are not valid since there is no ground truth depth, however in the visualization it looks like the confidence of the sky points is low. Could you please explain how you achieved low confidence for the sky points, are you using any additional loss for the invalid points?