Ci Hai ; Wu Mingdong ; Zhu Wentao ; Ma Xiaoxuan ; Dong Hao ; Zhong Fangwei ; Wang Yizhou

Abstract

Learning 3D human pose prior is essential to human-centered AI. Here, wepresent GFPose, a versatile framework to model plausible 3D human poses forvarious applications. At the core of GFPose is a time-dependent score network,which estimates the gradient on each body joint and progressively denoises theperturbed 3D human pose to match a given task specification. During thedenoising process, GFPose implicitly incorporates pose priors in gradients andunifies various discriminative and generative tasks in an elegant framework.Despite the simplicity, GFPose demonstrates great potential in severaldownstream tasks. Our experiments empirically show that 1) as amulti-hypothesis pose estimator, GFPose outperforms existing SOTAs by 20% onHuman3.6M dataset. 2) as a single-hypothesis pose estimator, GFPose achievescomparable results to deterministic SOTAs, even with a vanilla backbone. 3)GFPose is able to produce diverse and realistic samples in pose denoising,completion and generation tasks. Project pagehttps://sites.google.com/view/gfpose/