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:20260817T171531Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241203T131400
DTEND;TZID=Asia/Tokyo:20241203T132800
UID:siggraphasia_SIGGRAPH Asia 2024_sess104_papers_1191@linklings.com
SUMMARY:Taming 3DGS: High-Quality Radiance Fields with Limited Resources
DESCRIPTION:Saswat Subhajyoti Mallick (Carnegie Mellon University); Rahul 
 Goel (International Institute of Information Technology, Hyderabad); Bernh
 ard Kerbl (Carnegie Mellon University); Markus Steinberger (Graz Universit
 y of Technology); and Francisco Vicente Carrasco and Fernando De La Torre 
 (Carnegie Mellon University)\n\n3D Gaussian Splatting (3DGS) has transform
 ed novel-view synthesis with its fast, interpretable, and high-fidelity re
 ndering. However, its resource requirements limit its usability: Especiall
 y on weaker or constrained devices, training performance degrades quickly 
 and often cannot complete due to excessive memory consumption of the model
 . The method converges with an indefinite number of Gaussians---many of th
 em redundant---making rendering unnecessarily slow and preventing its usag
 e in downstream tasks that expect fixed-size inputs.\nTo address these iss
 ues, we tackle the challenges of training and rendering 3DGS models at a b
 udget. We use a guided, purely constructive densification process that ste
 ers densification to Gaussians that raise the reconstruction quality. Mode
 l size continuously increases in a controlled manner towards an exact budg
 et, using score-based densification of Gaussians with training-time priors
  that measure their contribution. We further address training speed obstac
 les: following a careful analysis of 3DGS' original pipeline, we derive fa
 ster, numerically equivalent solutions for gradient computation and attrib
 ute updates, including an alternative parallelization for efficient backpr
 opagation. We also propose quality-preserving approximations where suitabl
 e to reduce training time even further. Taken together, these enhancements
  yield a robust, scalable solution with reduced training times, lower comp
 ute and memory requirements, and high quality. Our evaluation shows that i
 n a budgeted setting, we obtain competitive quality metrics with 3DGS whil
 e achieving more than a 5x reduction in both model size and training time.
  With more generous budgets, our measured quality surpasses theirs. These 
 advances open the door for novel-view synthesis in constrained environment
 s, e.g., mobile or networked devices.\n\nRegistration Category: Full Acces
 s, Full Access Supporter\n\nLanguage Format: English Language\n\nSession C
 hair: Bernhard Kerbl (Technical University of Vienna)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_1191&sess=sess104
END:VEVENT
END:VCALENDAR