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:20250110T023312Z
LOCATION:Hall B5 (1)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241205T134600
DTEND;TZID=Asia/Tokyo:20241205T135800
UID:siggraphasia_SIGGRAPH Asia 2024_sess130_papers_292@linklings.com
SUMMARY:Neural Laplacian Operator for 3D Point Clouds
DESCRIPTION:Technical Papers\n\nBo Pang, Zhongtian Zheng, Yilong Li, Guopi
 ng Wang, and Peng-Shuai Wang (Peking University)\n\nThe Laplacian operator
  holds a crucial role in 3D geometry processing, yet it is still challengi
 ng to define it on point clouds.\nPrevious works mainly focused on constru
 cting a local triangulation around each point to approximate the underlyin
 g manifold for defining the Laplacian operator, which may not be very robu
 st or accurate.\nIn contrast, we simply use the $K$-nearest neighbors (KNN
 ) graph constructed from the input point cloud and learn the Laplacian ope
 rator on the KNN graph with graph neural networks (GNNs).\nHowever, the gr
 ound-truth Laplacian operator is defined on a manifold mesh with a differe
 nt connectivity from the KNN graph and thus cannot be directly used for tr
 aining.\nTo train the GNN, we propose a novel training scheme by imitating
  the behavior of the ground-truth Laplacian operator on a set of probe fun
 ctions so that the learned Laplacian operator behaves similarly to the gro
 und-truth Laplacian operator.\nWe train our network on a subset of ShapeNe
 t and evaluate it across a variety of point clouds.\nCompared with previou
 s methods, our method reduces the error by \emph{an order of magnitude} an
 d excels in handling sparse point clouds with thin structures or sharp fea
 tures.\nOur method also demonstrates a strong generalization ability to un
 seen shapes.\nWith our learned Laplacian operator, we further apply a seri
 es of Laplacian-based geometry processing algorithms directly to point clo
 uds and achieve accurate results, enabling many exciting possibilities for
  geometry processing on point clouds.\n\emph{We will release our code and 
 trained models to ensure reproducibility.}\n\nRegistration Category: Full 
 Access, Full Access Supporter\n\nLanguage Format: English Language\n\nSess
 ion Chair: Noam Aigerman (University of Montreal)
URL:https://asia.siggraph.org/2024/program/?id=papers_292&sess=sess130
END:VEVENT
END:VCALENDAR