Probabilistic 3D Human Shape and Pose Estimation from Multiple Unconstrained Images in the Wild

This paper addresses the problem of 3D human body shape and pose estimation\nfrom RGB images. Recent progress in this field has focused on single images,\nvideo or multi-view images as inputs. In contrast, we propose a new task: shape\nand pose estimation from a group of multiple images of a human subject, without\nconstraints on subject pose, camera viewpoint or background conditions between\nimages in the group. Our solution to this task predicts distributions over SMPL\nbody shape and pose parameters conditioned on the input images in the group. We\nprobabilistically combine predicted body shape distributions from each image to\nobtain a final multi-image shape prediction. We show that the additional body\nshape information present in multi-image input groups improves 3D human shape\nestimation metrics compared to single-image inputs on the SSP-3D dataset and a\nprivate dataset of tape-measured humans. In addition, predicting distributions\nover 3D bodies allows us to quantify pose prediction uncertainty, which is\nuseful when faced with challenging input images with significant occlusion. Our\nmethod demonstrates meaningful pose uncertainty on the 3DPW dataset and is\ncompetitive with the state-of-the-art in terms of pose estimation metrics.\n