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:20250110T023313Z
LOCATION:Hall B5 (1)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241206T135600
DTEND;TZID=Asia/Tokyo:20241206T141000
UID:siggraphasia_SIGGRAPH Asia 2024_sess145_papers_1060@linklings.com
SUMMARY:Multi-level Partition of Unity on Differentiable Moving Particles
DESCRIPTION:Technical Papers\n\nJinjin He and Taiyuan Zhang (Dartmouth Col
 lege); Hiroki Kobayashi and Atsushi Kawamoto (Toyota Central R&D Labs., In
 c.); Yuqing Zhou (Toyota Research Institute of North America); Tsuyoshi No
 mura (Toyota Central R&D Labs., Inc.); and Bo Zhu (Georgia Institute of Te
 chnology)\n\nWe introduce a differentiable moving particle representation 
 based on the\nmulti-level partition of unity (MPU) to represent dynamic im
 plicit geome-\ntries. At the core of our representation are two groups of 
 particles, named\nfeature particles and sample particles, which can move i
 n space and produce\ndynamic surfaces according to external velocity field
 s or optimization gradi-\nents. These two particle groups iteratively guid
 e and correct each other by\nalternating their roles as inputs and outputs
 . Each feature particle carries\na set of coefficients for a local quadrat
 ic patch. These particle patches are\nassembled with partition-of-unity we
 ights to derive a continuous implicit\nglobal shape. Each sampling particl
 e carries its position and orientation,\nserving as dense surface samples 
 for optimization tasks. Based on these mov-\ning particles, we develop a f
 ully differentiable framework to infer and evolve\nhighly detailed implici
 t geometries, enhanced by a multi-level background\ngrid for particle adap
 tivity, across different inverse tasks. We demonstrated\nthe efficacy of o
 ur representation through various benchmark comparisons\nwith state-of-the
 -art neural representations, achieving lower memory con-\nsumption, fewer 
 training iterations, and orders of magnitude higher accuracy\nin handling 
 topologically complex objects and dynamic tracking tasks.\n\nRegistration 
 Category: Full Access, Full Access Supporter\n\nLanguage Format: English L
 anguage\n\nSession Chair: Hao (Richard) Zhang (Simon Fraser University, Am
 azon)
URL:https://asia.siggraph.org/2024/program/?id=papers_1060&sess=sess145
END:VEVENT
END:VCALENDAR