PaperReading

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

• Uncertainty by image: 2 forward pass
  • flipping: easiest one.
• Empirical estimation: N forward pass
  • Dropout sampling: turn on random dropout during prediction
  • Bootstrapped Ensemble: same model, N diff initialization
  • Snapshot Ensemble: same model, N early stopped version with cyclic LR
• Predictive estimation: 1 forward pass
  • Learned reprojection: anomaly prediction predicting the
  • Log-likelihood Maximization: whatever loss + aleatoric uncertainty. –> Learn SfM from SfM
  • Self-teaching: L1/L2 loss + aleatoric uncertainty
• Bayesian estimation
  • Combination of empirical and predictive estimation. Uncertainty is the sum of predicted uncertainty, plus the deviation of predicted depth from the averaged depth.
• The conclusions:
  • Bootstrap ensemble is about the same as snapshot ensemble, but slightly better.
  • For monocular setup, empirical methods does not work well. Self-teaching improves baseline while log-likelihood worsens baseline.
  • For uncertainty, self-teaching > log-likelihood > postprocessing.

Technical details

• Weakly supervised:
  • Noisy lidar depth
  • Model-based depth: SGM for stereo, SfM, and with their confidence
• Uncertainty metric:
  • Sparsification plot (see explanation in Active learning in Lidar) to compare with oracle sparsification. It is useful to compare each model with its oracle. It is usually normalized to 1 at 0 removal (max error).
  • Sparsification error is the difference between each model and its oracle sparsification. It is therefore possible to compare different models. It by definition starts at [0, 0].
  • The single metric summary of sparsification error is AUSE (area under the sparsification error)
  • The idea was first introduced in Uncertainty Estimates and Multi-Hypotheses Networks for Optical Flow ECCV 2018

Notes

• Questions and notes on how to improve/revise the current work