ICCV Daily 2021 - Friday

The key takeaway message from the paper is that by taking this one step further and showing that this learned distribution is indeed useful for a variety of downstream tasks, it can combine information from multiple sources and generate more accurate predictions. “ The most difficult part of the work has been using these probabilistic models, normalizing flows for instance, in practice, ” Nikos tells us. “ We were very familiar with the human pose estimation task, but the challenge was to incorporate this probabilistic model in there. It’s different to read about them and what they do in principle. Making them work in practice takes a lot of experimentation. ” “ I totally agree with Nikos, ” Georgios adds. “ The problem of human mesh recovery is something that we are quite familiar with. We had some previous works recently, so we knew what was hard and what was interesting to do in that direction. However, using this toolbox of normalizing flows and all the advantages that it has, while it was very appealing to us as an idea, it was really hard to integrate properly. All credit goes to Nikos for making this work so nicely. ” 13 DAILY ICCV Friday Nikos Kolotouros and Georgios Pavlakos