Question on the depth supervision

[jizongFox]

Hello,

Thank you for this great work. I have a small question regarding the depth theory used in the paper. I am aware that there is shift and scale invariant way to train mono-depth models, such as DPT and more recently depth anything. The ambiguity here is that they applied this principle in disparity space, which is the inverse of the true depth value.

Therefore, using a model trained with scale and shift invariant principle on disparity to generate a depth value, and then, apply this scale and shift invariant principle again on the depth space may cause some confusion.

Quite new to this and what's your thought?

Best,
Jizong Peng