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:20260817T171530Z
LOCATION:Hall B5 (2)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241204T111900
DTEND;TZID=Asia/Tokyo:20241204T113100
UID:siggraphasia_SIGGRAPH Asia 2024_sess113_papers_1022@linklings.com
SUMMARY:Dynamic Gaussian Marbles for Novel View Synthesis of Casual Monocu
 lar Videos
DESCRIPTION:Colton Stearns, Adam Harley, and Mikaela Uy (Stanford Universi
 ty); Florian Dubost and Federico Tombari (Google Research); and Gordon Wet
 zstein and Leonidas Guibas (Stanford University)\n\nGaussian splatting has
  become a popular representation for novel-view synthesis, exhibiting clea
 r strengths in efficiency, photometric quality, and compositional edibilit
 y. Following its success, many works have extended Gaussians to 4D, showin
 g that dynamic Gaussians maintain these benefits while also tracking scene
  geometry far better than alternative representations. Yet, these methods 
 assume dense multi-view videos as supervision, constraining their use to c
 ontrolled capture settings. In this work, we are interested in extending t
 he capability of Gaussian scene representations to casually captured monoc
 ular videos. We show that existing 4D Gaussian methods dramatically fail i
 n this setup because the monocular setting is underconstrained. Building o
 ff this finding, we propose a method we call Dynamic Gaussian Marbles, whi
 ch consist of three core modifications that target the difficulties of the
  monocular setting. First, we use isotropic Gaussian "marbles", reducing t
 he degrees of freedom of each Gaussian, and constraining the optimization 
 to focus on motion and appearance over local shape. Second, we employ a hi
 erarchical divide-and-conquer learning strategy to efficiently guide the o
 ptimization towards solutions with globally coherent motion. Finally, we a
 dd image-level and geometry-level priors into the optimization, including 
 a tracking loss that takes advantage of recent progress in point tracking.
  By constraining the optimization in these ways, Dynamic Gaussian Marbles 
 learns Gaussian trajectories that enable novel-view rendering and accurate
 ly capture the 3D motion of the scene elements. We evaluate on the (monocu
 lar) Nvidia Dynamic Scenes dataset and the Dycheck iPhone dataset, and sho
 w that Gaussian Marbles significantly outperforms other Gaussian baselines
  in quality, and is on-par with non-Gaussian representations, all while ma
 intaining the efficiency, compositionality, editability, and tracking bene
 fits of Gaussians.\n\nRegistration Category: Full Access, Full Access Supp
 orter\n\nLanguage Format: English Language\n\nSession Chair: Forrester Col
 e (Google)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_1022&sess=sess113
END:VEVENT
END:VCALENDAR