This repository provides the code for the ICCV'19 publication "Sampling-free Epistemic Uncertainty Estimation Using Approximated Variance Propagation". We provide a sampling-free approach for estimating epistemic uncertainty when applying methods based on noise injection (e.g. stochastic regularization). Our approach is motivated by error propagation. We primarily compare our approach with Monte-Carlo (MC) dropout by approximating the sampling procdeure of the latter.
Following the experiment section in our paper, this repository is divided into three sections:
- UCI_regression: Comparison of predictive performance between MC dropout and our variance propagation approach. The code further implements Mixture Density Networks and learning the dropout parameter using our sampling-free approximation, which is not included in our publication.
- Bayesian_SegNet: Sampling-free approximation of Bayesian SegNet. We investigate the quality of our approximation on a high-dimensional semantic segmentation task.
- Monocular_Depth_Regression: Applies our approximation to Unsupervised Monocular Depth Regression. We enhance the original work by inserting dropout at the final layers and analyze the quality of our approximation.