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:20250110T023309Z
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:Technical Papers\n\nSaswat Subhajyoti Mallick (Carnegie Mellon
  University); Rahul Goel (International Institute of Information Technolog
 y, Hyderabad); Bernhard Kerbl (Carnegie Mellon University); Markus Steinbe
 rger (Graz University of Technology); and Francisco Vicente Carrasco and F
 ernando De La Torre (Carnegie Mellon University)\n\n3D Gaussian Splatting 
 (3DGS) has transformed novel-view synthesis with its fast, interpretable, 
 and high-fidelity rendering. However, its resource requirements limit its 
 usability: Especially on weaker or constrained devices, training performan
 ce degrades quickly and often cannot complete due to excessive memory cons
 umption of the model. The method converges with an indefinite number of Ga
 ussians---many of them redundant---making rendering unnecessarily slow and
  preventing its usage in downstream tasks that expect fixed-size inputs.\n
 To address these issues, we tackle the challenges of training and renderin
 g 3DGS models at a budget. We use a guided, purely constructive densificat
 ion process that steers densification to Gaussians that raise the reconstr
 uction quality. Model size continuously increases in a controlled manner t
 owards an exact budget, using score-based densification of Gaussians with 
 training-time priors that measure their contribution. We further address t
 raining speed obstacles: following a careful analysis of 3DGS' original pi
 peline, we derive faster, numerically equivalent solutions for gradient co
 mputation and attribute updates, including an alternative parallelization 
 for efficient backpropagation. We also propose quality-preserving approxim
 ations where suitable to reduce training time even further. Taken together
 , these enhancements yield a robust, scalable solution with reduced traini
 ng times, lower compute and memory requirements, and high quality. Our eva
 luation shows that in a budgeted setting, we obtain competitive quality me
 trics with 3DGS while achieving more than a 5x reduction in both model siz
 e and training time. With more generous budgets, our measured quality surp
 asses theirs. These advances open the door for novel-view synthesis in con
 strained environments, e.g., mobile or networked devices.\n\nRegistration 
 Category: Full Access, Full Access Supporter\n\nLanguage Format: English L
 anguage\n\nSession Chair: Bernhard Kerbl (Technical University of Vienna)
URL:https://asia.siggraph.org/2024/program/?id=papers_1191&sess=sess104
END:VEVENT
END:VCALENDAR