Most previous works treat human pose estimation from a single image as a typical regression problem, where given an image as input, they try to predict a single pose result. In this paper, Nikos and Georgios argue that this is suboptimal, since the problem of predicting a 3D pose from a single image is inherently ambiguous. Other works have considered moving away from single predictions, but usually they do it within a multiple hypothesis framework, where instead of a single output, they have multiple different outputs. Usually, the number of outputs is fixed in this case. This work does not output multiple hypotheses, but rather regresses a distribution of plausible poses given the input image. 12 DAILY ICCV Friday Poster Presentation Probabilistic Modeling for Human Mesh Recovery Georgios Pavlakos is a postdoc at UC Berkeley, working with Angjoo Kanazawa and Jitendra Malik. Nikos Kolotouros is a fifth- year PhD student at the University of Pennsylvania, advised by Kostas Daniilidis, and is currently a research intern at Facebook Reality Labs. Their work introduces a new paradigm for human pose estimation and has been accepted as a poster at ICCV this year. Nikos and Georgios speak to us ahead of their live Q&A session today.