Learning-Deep-Learning

MonoUncertainty: On the uncertainty of self-supervised monocular depth estimation

July 2020

tl;dr: Self-teaching with pseudo-label is the best way for uncertainty estimation for monocular depth estimation.

Overall impression

The paper has a very good review session on monocular depth estimation. Tons of ablations studies. It evaluated 11 methods and combinations for predicting the uncertainty of self-supervised monocular depth estimation.

Verdict is: When pose is known, the difference between self-teaching and log-likelihood is minor. When pose is entangled in the loss function, self-teaching is much better to produce the uncertainty of depth.

Key ideas

Technical details

Notes