Guo Wen ; Corona Enric ; Moreno-Noguer Francesc ; Alameda-Pineda Xavier

Abstract

Recent literature addressed the monocular 3D pose estimation task verysatisfactorily. In these studies, different persons are usually treated asindependent pose instances to estimate. However, in many every-day situations,people are interacting, and the pose of an individual depends on the pose ofhis/her interactees. In this paper, we investigate how to exploit thisdependency to enhance current - and possibly future - deep networks for 3Dmonocular pose estimation. Our pose interacting network, or PI-Net, inputs theinitial pose estimates of a variable number of interactees into a recurrentarchitecture used to refine the pose of the person-of-interest. Evaluating sucha method is challenging due to the limited availability of public annotatedmulti-person 3D human pose datasets. We demonstrate the effectiveness of ourmethod in the MuPoTS dataset, setting the new state-of-the-art on it.Qualitative results on other multi-person datasets (for which 3D poseground-truth is not available) showcase the proposed PI-Net. PI-Net isimplemented in PyTorch and the code will be made available upon acceptance ofthe paper.