Learning-Deep-Learning

BayesOD: A Bayesian Approach for Uncertainty Estimation in Deep Object Detectors

November 2019

tl;dr: Fuse aleatoric and epistemic uncertainty into bayesian framework.

Overall impression

Very statistical paper. Lots of math details. The implementation in tensorflow probability (tfp) package. It reinterprets the DL object detector as one measurement device, with variance.

The anchor boxes are different measurement devices.

Marginalization over anchor distribution seem to predict epistemic uncertainty

Open-set detection problem (detecting unknown unknowns) is still one area safety-critical applications such as autonomous driving needs to focus on.

Key ideas

Technical details

Notes