Distance to the Sky is mostly estimated <2km? Instruction appreciated for re-train/ fine-tune, thanks!

[lattebyte]

Greetings,

Please find 2 images below that were generated from a modified DepthPro in point-cloud-and-mapping, it is using original Apple DepthPro AI model,

inferred_1.png (looks no major issues):

road1_2d.png:

This "road1_2d.png" is the one that included sky where most of sky-pixel are estimated below 2000m, although max value is ~10km. however chatgpt suggests is 10-50km to Stratosphere.

It would be appreciated if anyone could please advise what is the theory/ referece documentation of how the Apple original model is trained (especially to the sky), and how to fine-tune or re-train.

I dont have much knowledge in depth-related AI model, but some experice with creating mask/ oject detection .pt or .pth model, thanks!

video demo:
https://youtu.be/3Vm2yJsiLZY