Su Jiajun ; Wang Chunyu ; Ma Xiaoxuan ; Zeng Wenjun ; Wang Yizhou

Abstract

While monocular 3D pose estimation seems to have achieved very accurateresults on the public datasets, their generalization ability is largelyoverlooked. In this work, we perform a systematic evaluation of the existingmethods and find that they get notably larger errors when tested on differentcameras, human poses and appearance. To address the problem, we introduceVirtualPose, a two-stage learning framework to exploit the hidden "free lunch"specific to this task, i.e. generating infinite number of poses and cameras fortraining models at no cost. To that end, the first stage transforms images toabstract geometry representations (AGR), and then the second maps them to 3Dposes. It addresses the generalization issue from two aspects: (1) the firststage can be trained on diverse 2D datasets to reduce the risk of over-fittingto limited appearance; (2) the second stage can be trained on diverse AGRsynthesized from a large number of virtual cameras and poses. It outperformsthe SOTA methods without using any paired images and 3D poses from thebenchmarks, which paves the way for practical applications. Code is availableat https://github.com/wkom/VirtualPose.