BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Tokyo
X-LIC-LOCATION:Asia/Tokyo
BEGIN:STANDARD
TZOFFSETFROM:+0900
TZOFFSETTO:+0900
TZNAME:JST
DTSTART:18871231T000000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260817T171532Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241203T132800
DTEND;TZID=Asia/Tokyo:20241203T134200
UID:siggraphasia_SIGGRAPH Asia 2024_sess104_papers_415@linklings.com
SUMMARY:Representing Long Volumetric Video with Temporal Gaussian Hierarch
 y
DESCRIPTION:Zhen Xu (State Key Laboratory of CAD&CG, Zhejiang University; 
 Zhejiang University); Yinghao Xu (Stanford University); Zhiyuan Yu (Depart
 ment of Mathematics, Hong Kong University of Science and Technology); Sida
  Peng and Jiaming Sun (Zhejiang University); and Hujun Bao and Xiaowei Zho
 u (State Key Laboratory of CAD&CG, Zhejiang University)\n\nThis paper aims
  to address the challenge of reconstructing long volumetric videos from mu
 lti-view RGB videos.\nRecent dynamic view synthesis methods leverage power
 ful 4D representations, like feature grids or point cloud sequences, to ac
 hieve high-quality rendering results. However, they are typically limited 
 to short (1$\sim$2s) video clips and often suffer from large memory footpr
 ints when dealing with longer videos.\nTo solve this issue, we propose a n
 ovel 4D representation, named temporal Gaussian hierarchy, to compactly mo
 del long volumetric videos.\nOur key observation is that there are general
 ly various degrees of temporal redundancy in dynamic scenes, which consist
  of areas changing at different speeds.\nMotivated by this, our approach b
 uilds a multi-level hierarchy of Gaussian primitives, where each level sep
 arately describes scene regions with different degrees of content change, 
 and adaptively shares Gaussian primitives to represent unchanged scene con
 tent over different temporal segments, thus effectively reducing the numbe
 r of Gaussian primitives.\nIn addition, the tree-like structure of the Gau
 ssian hierarchy allows us to efficiently represent the scene at a particul
 ar moment with a subset of Gaussian primitives, leading to nearly constant
  GPU memory usage during the training or rendering regardless of the video
  length.\nMoreover, we design a compact appearance model that mixes diffus
 e and view-dependent Gaussians to further minimize the model size while ma
 intaining the rendering quality.\nWe also develop a rasterization pipeline
  of Gaussian primitives based on the hardware-accelerated technique to imp
 rove rendering speed.\nExtensive experimental results demonstrate the supe
 riority of our method over alternative methods in terms of training cost, 
 rendering speed, and storage usage.\nTo our knowledge, this work is the fi
 rst approach capable of efficiently handling hours of volumetric video dat
 a while maintaining state-of-the-art rendering quality.\n\nRegistration Ca
 tegory: Full Access, Full Access Supporter\n\nLanguage Format: English Lan
 guage\n\nSession Chair: Bernhard Kerbl (Technical University of Vienna)\n\
 n
URL:https://asia.siggraph.org/2024/program/?id=papers_415&sess=sess104
END:VEVENT
END:VCALENDAR