BEGIN:VCALENDAR
VERSION:2.0
PRODID:Linklings LLC
BEGIN:VTIMEZONE
TZID:Asia/Hong_Kong
X-LIC-LOCATION:Asia/Hong_Kong
BEGIN:STANDARD
TZOFFSETFROM:+0800
TZOFFSETTO:+0800
TZNAME:HKT
DTSTART:19911015T033000
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20251218T030655Z
LOCATION:Meeting Room S423+S424\, Level 4
DTSTART;TZID=Asia/Hong_Kong:20251217T170200
DTEND;TZID=Asia/Hong_Kong:20251217T171300
UID:siggraphasia_SIGGRAPH Asia 2025_sess144_papers_1794@linklings.com
SUMMARY:Robust Single-shot Structured Light 3D Imaging via Neural Feature 
 Decoding
DESCRIPTION:Jiaheng Li (Wangxuan Institute of Computer Technology, Peking 
 University); Qiyu Dai (School of Intelligence Science and Technology, Peki
 ng University); Lihan Li (Yuanpei College, Peking University); Praneeth Ch
 akravarthula (University of North Carolina at Chapel Hill (UNC)); He Sun (
 College of Future Technology, Peking University); Baoquan Chen (School of 
 Intelligence Science and Technology, Peking University; State Key Laborato
 ry of General Artificial Intelligence, Peking University); and Wenzheng Ch
 en (Wangxuan Institute of Computer Technology, Peking University; State Ke
 y Laboratory of General Artificial Intelligence, Peking University)\n\nWe 
 consider the problem of active 3D imaging using single-shot structured lig
 ht systems, which are widely employed in commercial 3D sensing devices suc
 h as Apple Face ID and Intel RealSense. Traditional structured light metho
 ds typically decode depth correspondences through pixel-domain matching al
 gorithms, resulting in limited robustness under challenging scenarios like
  occlusions, fine-structured details, and non-Lambertian surfaces. Inspire
 d by recent advances in neural feature matching, we propose a learning-bas
 ed structured light decoding framework that performs robust correspondence
  matching within feature space rather than the fragile pixel domain. Our m
 ethod extracts neural features from the projected patterns and captured in
 frared (IR) images, explicitly incorporating their geometric priors by bui
 lding cost volumes in feature space, achieving substantial performance imp
 rovements over pixel-domain decoding approaches. To further enhance depth 
 quality, we introduce a depth refinement module that leverages strong prio
 rs from large-scale monocular depth estimation models, improving fine deta
 il recovery and global structural coherence. To facilitate effective learn
 ing, we develop a physically-based structured light rendering pipeline, ge
 nerating nearly one million synthetic pattern-image pairs with diverse obj
 ects and materials for indoor settings. Experiments demonstrate that our m
 ethod, trained exclusively on synthetic data with multiple structured ligh
 t patterns, generalizes well to real-world indoor environments, effectivel
 y processes various pattern types without retraining, and consistently out
 performs both commercial structured light systems and passive stereo RGB-b
 ased depth estimation methods.\n\nRegistration Category: Full Access, Full
  Access Supporter\n\nSession Chair: Seung-Hwan Baek (POSTECH)\n\n
END:VEVENT
END:VCALENDAR