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 estimationfrom RGB images. Recent progress in this field has focused on single images,video or multi-view images as inputs. In contrast, we propose a new task: shapeand pose estimation from a group of multiple images of a human subject, withoutconstraints on subject pose, camera viewpoint or background conditions betweenimages in the group. Our solution to this task predicts distributions over SMPLbody shape and pose parameters conditioned on the input images in the group. Weprobabilistically combine predicted body shape distributions from each image toobtain a final multi-image shape prediction. We show that the additional bodyshape information present in multi-image input groups improves 3D human shapeestimation metrics compared to single-image inputs on the SSP-3D dataset and aprivate dataset of tape-measured humans. In addition, predicting distributionsover 3D bodies allows us to quantify pose prediction uncertainty, which isuseful when faced with challenging input images with significant occlusion. Ourmethod demonstrates meaningful pose uncertainty on the 3DPW dataset and iscompetitive with the state-of-the-art in terms of pose estimation metrics.