Thank you for your great work and releasing UnidepthV2!
I have a question regarding the “confidence” / “uncertainty” output.
In the paper (e.g., Fig. 5), you mention that confidence is obtained as the inverse uncertainty. However, in the current implementation it looks like the decoder outputs an uncertainty map, and this tensor named "confidence" in this codebase without an explicit inverse transform.
From reading the code, my understanding is that the network predicts a per-pixel uncertainty (approximate log-depth error), and higher values mean lower confidence, but the inverse (to obtain a proper “confidence” map) is not applied inside the model.
Could you please clarify the intended behavior?
Thank you for your great work and releasing UnidepthV2!
I have a question regarding the “confidence” / “uncertainty” output.
In the paper (e.g., Fig. 5), you mention that confidence is obtained as the inverse uncertainty. However, in the current implementation it looks like the decoder outputs an uncertainty map, and this tensor named "confidence" in this codebase without an explicit inverse transform.
From reading the code, my understanding is that the network predicts a per-pixel uncertainty (approximate log-depth error), and higher values mean lower confidence, but the inverse (to obtain a proper “confidence” map) is not applied inside the model.
Could you please clarify the intended behavior?