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:20260817T171532Z
LOCATION:Hall B5 (1)\, B Block\, Level 5
DTSTART;TZID=Asia/Tokyo:20241204T144500
DTEND;TZID=Asia/Tokyo:20241204T145900
UID:siggraphasia_SIGGRAPH Asia 2024_sess118_papers_431@linklings.com
SUMMARY:GarVerseLOD: High-Fidelity 3D Garment Reconstruction from a Single
  In-the-Wild Image using a Dataset with Levels of Details
DESCRIPTION:Zhongjin Luo, Haolin Liu, Chenghong Li, Wanghao Du, Zirong Jin
 , and Wanhu Sun (Chinese University of Hong Kong, Shenzhen); Yinyu Nie (Hu
 awei Technologies Ltd.); Weikai Chen (Tencent America); and Xiaoguang Han 
 (Chinese University of Hong Kong, Shenzhen)\n\nNeural implicit functions h
 ave brought impressive advances to the state-of-the-art of clothed human d
 igitization from multiple or even single images. However, despite the prog
 ress, current arts still have difficulty generalizing to unseen images wit
 h complex cloth deformation and body poses. In this work, we present GarVe
 rseLOD, a new dataset and framework that paves the way to achieving unprec
 edented robustness in high-fidelity 3D garment reconstruction from a singl
 e unconstrained image. Inspired by the recent success of large generative 
 models, we believe that one key to addressing the generalization challenge
  lies in the quantity and quality of 3D garment data. Towards this end, Ga
 rVerseLOD collects 6,000 high-quality cloth models with fine-grained geome
 try details manually created by professional artists. In addition to the s
 cale of training data, we observe that having disentangled granularities o
 f geometry can play an important role in boosting the generalization capab
 ility and inference accuracy of the learned model.	We hence craft GarVerse
 LOD as a hierarchical dataset with levels of details (LOD), spanning from 
 detail-free stylized shape to pose-blended garment with pixel-aligned deta
 ils. This allows us to make this highly under-constrained problem tractabl
 e by factorizing the inference into easier tasks, each narrowed down with 
 smaller searching space. To ensure GarVerseLOD can generalize well to in-t
 he-wild images, we propose a novel labeling paradigm based on conditional 
 diffusion models to generate extensive paired images for each garment mode
 l with high photorealism. We evaluate our method on a massive amount of in
 -the-wild images. Experimental results demonstrate that GarVerseLOD can ge
 nerate standalone garment pieces with significantly better quality than pr
 ior approaches while being robust against a large variation of pose, illum
 ination, occlusion, and deformation. Code and dataset are available at gar
 verselod.github.io.\n\nRegistration Category: Full Access, Full Access Sup
 porter\n\nLanguage Format: English Language\n\nSession Chair: Meng Zhang (
 Nanjing University of Science and Technology)\n\n
URL:https://asia.siggraph.org/2024/program/?id=papers_431&sess=sess118
END:VEVENT
END:VCALENDAR