Ilg E, Cicek O, Galesso S, et al. Uncertainty estimates and multi-hypotheses networks for optical flow[C]//Proceedings of the European Conference on Computer Vision (ECCV). 2018: 652-667.
1. Overview
1.1. Motivation
- it is not possible to deploy such a system without information about how reliable the underlying estimates are
- we should expect an additional estimate of the network’s own uncertainty, such that the network can highlight hard cases where it cannot reliably estimate
In this paper
- compare several strategies and techniques to estimate uncertainty
- introduce a network utilizing the Winner-Takes-All (WTA, penalize only the best prediction) loss without the need for sampling or ensembles
1.2. Network
- (e). the proposed network. generate multi-hypothesis (distribution) then merge