Test-time Personalizable Forecasting of 3D Human Poses

Paper accepted by ICCV 2023

Qiongjie Cui, Huaijiang Sun, Jianfeng Lu, Weiqing Li, Bin Li, Hongwei Yi, Haofan Wang

We propose a new task: How to perform high-fidelity predictions when there are distribution gaps between testing and training characters ?

Demo Video and Visulazition of Our Method

(Eating activity) Predicted poses of SOTA method (ECCV23) and ours+SPGSN

SPGSN (ECCV23)

MPJPE (at end pose)=73.4mm









SPGSN w/ H/PTTP

MPJPE (at end pose)=67.3mm:



(Smoking activity) Predicted poses of SOTA method (ECCV23) and ours+SPGSN

SPGSN (ECCV23)

MPJPE (at end pose)=68.6mm









SPGSN w/ H/PTTP

MPJPE (at end pose)=63.7mm


(Discussion activity) Predicted poses of SOTA method (ECCV23) and ours+SPGSN

SPGSN (ECCV23)

MPJPE (at end pose)=116.8mm










SPGSN w/ H/PTTP

MPJPE (at end pose)=112.5mm


Cite this paper using the following context:

@inproceedings{cui2023test,

  title={Test-time Personalizable Forecasting of 3D Human Poses},

  author={Qiongjie Cui, Huaijiang Sun, Jianfeng Lu, Weiqing Li, Bin Li, Hongwei Yi, Haofan Wang},

  booktitle={ICCV},

  year={2023}

}

The camera-ready paper is available at: https://drive.google.com/file/d/1LWtw-JmSqRodN1fNSgGL6wUdOEKw5uWu/view?usp=drive_link

Code will be publicly released.

This work is partly supported by Xiaohongshu Inc. China