Bayesian SegNet: Model Uncertainty in Deep Convolutional Encoder-Decoder Architectures for Scene Understanding

June 2019

tl;dr: Estimate the variance of segmentation uncertainty with dropout inference samples. Use the mean for prediction. The idea is quite similar to TTA (test time augmentation).

Overall impression

The paper provides a practical way to evaluate the uncertainty (this is the epistemic uncertainty), at a cost at inference time. Refer to Bayesian DL for integration with aleatoric uncertainty.

Key ideas

Technical details