BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030653Z
LOCATION:Meeting Room S421\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251216T154400
DTEND;TZID=Asia/Hong_Kong:20251216T155500
UID:siggraphasia_SIGGRAPH Asia 2025_sess124_papers_1986@linklings.com
SUMMARY:Split4D: Decomposed 4D Scene Reconstruction Without Video Segmenta
 tion
DESCRIPTION:Yongzhen Hu (State Key Laboratory of CAD&CG, Zhejiang Universi
 ty; Ant Group); Yihui Yang, Haotong Lin, and Yifan Wang (State Key Laborat
 ory of CAD&CG, Zhejiang University); Junting Dong (Shanghai Artificial Int
 elligence Laboratory); Yifu Deng, Xinyu Zhu, and Fan Jia (Ant Group); and 
 Hujun Bao, Xiaowei Zhou, and Sida Peng (State Key Laboratory of CAD&CG, Zh
 ejiang University)\n\nThis paper addresses the problem of decomposed 4D sc
 ene reconstruction from multi-view videos.\nRecent methods achieve this by
  lifting video segmentation results to a 4D representation through differe
 ntiable rendering techniques.\nTherefore, they heavily rely on the quality
  of video segmentation maps, which are often unstable, leading to unreliab
 le reconstruction results.\nTo overcome this challenge, our key idea is to
  represent the decomposed 4D scene with the Freetime FeatureGS and design 
 a streaming feature learning strategy to accurately recover it from per-im
 age segmentation maps, eliminating the need for video segmentation.\nFreet
 ime FeatureGS models the dynamic scene as a set of Gaussian primitives wit
 h learnable features and linear motion ability, allowing them to move to n
 eighboring regions over time.\nWe apply a contrastive loss to Freetime Fea
 tureGS, forcing primitive features to be close or far apart based on wheth
 er their projections belong to the same instance in the 2D segmentation ma
 p.\nAs our Gaussian primitives can move across time, it naturally extends 
 the feature learning to the temporal dimension, achieving 4D segmentation.
 \nFurthermore, we sample observations for training in a temporally ordered
  manner, enabling the streaming propagation of features over time and effe
 ctively avoiding local minima during the optimization process.\nExperiment
 al results on several datasets show that the reconstruction quality of our
  method outperforms recent methods by a large margin.\n\nRegistration Cate
 gory: Full Access, Full Access Supporter\n\nSession Chair: Yan-Pei Cao (VA
 ST)\n\n
END:VEVENT
END:VCALENDAR