Question Regarding Sky Handling During Training

[rainyNighti]

Hello,
I have a question about how the model handles the sky during the training process.
In Depth Anything v2, the model predicts disparity. It uses a segmentation model to mark the sky and sets its ground truth (GT) disparity to zero.

Since this new work predicts affine-invariant depth instead of disparity, I was wondering how you handle the sky. Specifically:

• Do you set a specific ground truth value for the sky, such as zero or a maximum depth value?
• Do you still use a segmentation model, like SegFormer, to identify the sky?

Thank you for your time and clarification!