notesum.ai

Published at November 22

Morph: A Motion-free Physics Optimization Framework for Human Motion Generation

cs.CV

Released Date: November 22, 2024

Authors: Zhuo Li¹, Mingshuang Luo², Ruibing Hou², Xin Zhao³, Hao Liu¹, Hong Chang², Zimo Liu⁴, Chen Li¹

Aff.: ¹WeChat, Tencent Inc; ²Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, China; ³MoE Key Laboratory of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University; ⁴Peng Cheng Laboratory, China

Arxiv: http://arxiv.org/abs/2411.14951v1

Model Number	Methods					Common Generation Metrics		Physical Plausibility Metrics
Model Number	IS	Adversarial	Energy	Real Data	FT	RTOP-3 $\uparrow$	FID $\downarrow$	PFC $\downarrow$	Penetrate $\downarrow$	Float $\downarrow$	Skate $\downarrow$	IFR $\downarrow$
$\mathrm{A}$	Baseline (only motion generator MoMask [7])					0.807	0.045	1.058	23.152	10.660	5.262	-
$\mathrm{B}$		✓	✓			0.792	0.194	0.852	0.000	2.272	0.020	0.0155
$\mathrm{C}$	✓		✓			0.782	0.276	0.735	0.000	2.554	0.032	0.0362
$\mathrm{D}$	✓	✓				0.790	0.165	0.715	0.000	2.376	0.026	0.0272
$\mathrm{E}$	✓	✓	✓			0.801	0.077	0.672	0.000	2.271	0.011	0.0155
$\mathrm{F}$	✓	✓	✓	✓		0.785	0.183	0.749	0.000	2.451	0.017	0.0338
$\mathrm{G}$	✓	✓	✓		✓	0.816	0.043	0.651	0.000	2.146	0.010	0.0151