Questions for "Hierarchical Discrete Distribution Decomposition for Match Density Estimation"

-----------------------------------------------------------------------------------------------
Please send your answers to: schroepp@cs.uni-freiburg.de by 10:00 on 29.01.2020

1. The approach presented in the paper predicts matching distributions between two images in a stereo or optical flow problem setting. Informally describe what is meant by matching distributions and how they differ from predictions of typical optical flow / stereo approaches? (2 sentences)

2. Due to the high computational cost of estimating full match distributions, the authors suggest decomposing the match distributions hierarchically into local match distributions. Give a short overview of how the estimation of this hierarchically decomposed distribution is actually done? (3 sentences)

3. Which loss is used to train the network and in what form does it use the given ground truth flow / disparities? (2 sentences)