notesum.ai

Published at November 26

cs.CV

Released Date: November 26, 2024

Authors: Ziyi Xu¹, Ziyao Huang¹, Juan Cao¹, Yong Zhang², Xiaodong Cun³, Qing Shuai⁴, Yuchen Wang¹, Linchao Bao⁴, Jintao Li¹, Fan Tang¹

Aff.: ¹Institute of Computing Technology, Chinese Academy of Sciences; ²Meituan; ³Great Bay University; ⁴Tencent

Method	FID↓	FVD↓	FID-VID↓	Obj-IoU↑	Obj-CLIP↑	LMD (Hand)↓	LMD (Body)↓	Subj-Cons↑	Back-Cons↑
AnyV2V	234.1	2873.3	53.1	0.241	0.744	94.6	30.5	68.2	77.9
MimicMotion+AnyDoor	167.8	1668.9	37.7	0.647	0.863	13.2	26.0	93.4	90.7
AnimateAnyone	172.8	2267.0	24.2	0.361	0.832	23.2	41.5	94.9	94.1
MimicMotion	138.1	1444.9	22.3	0.411	0.876	12.1	24.3	96.3	93.9
Ours	141.7	736.5	15.0	0.848	0.919	11.7	25.7	97.4	95.3
w/o Human-Object Dual Adapter	170.5	913.0	24.2	0.802	0.886	12.1	20.9	96.4	94.9
w/o Multi-View Object Feature Fusion	177.0	1371.9	52.0	0.845	0.908	12.0	24.6	97.2	95.3
w/o 3D Hand Mesh	164.0	920.1	21.3	0.847	0.907	12.0	22.3	97.3	95.0
w/o HOI-Region Reweighting Loss	164.7	807.3	69.9	0.846	0.895	11.8	22.0	97.0	94.8