GarVerseLOD
收藏资源简介:
GarVerseLOD是由中国香港中文大学(深圳)创建的高保真3D服装重建数据集,包含6000个由专业艺术家手工制作的精细几何细节的服装模型。数据集分为三个层次的细节级别(LOD),从无细节的粗略形状到带有像素对齐细节的姿势混合服装。创建过程中,利用条件扩散模型生成大量高质量的配对图像,以增强数据集的泛化能力。该数据集主要应用于从单张野外图像中重建高保真3D服装,旨在解决现有方法在处理复杂服装变形和多样化姿态时的不足。
GarVerseLOD is a high-fidelity 3D clothing reconstruction dataset developed by The Chinese University of Hong Kong, Shenzhen. It contains 6000 clothing models with fine geometric details, manually crafted by professional artists. The dataset is structured into three levels of detail (LOD), ranging from coarse shapes with no details to pose-blended clothing equipped with pixel-aligned details. During the dataset construction phase, conditional diffusion models were employed to generate a large number of high-quality paired images, thereby enhancing the generalization ability of the dataset. This dataset is primarily intended for high-fidelity 3D clothing reconstruction from a single in-the-wild image, aiming to address the shortcomings of existing methods when dealing with complex clothing deformations and diverse poses.
GarVerseLOD: High-Fidelity 3D Garment Reconstruction from a Single In-the-Wild Image using a Dataset with Levels of Details
数据集概述
- 标题: GarVerseLOD: High-Fidelity 3D Garment Reconstruction from a Single In-the-Wild Image using a Dataset with Levels of Details
- 作者:
- Zhongjin Luo<sup>1</sup>
- Haolin Liu<sup>2,1</sup>
- Chenghong Li<sup>2, 1</sup>
- Wanghao Du<sup>2</sup>
- Zirong Jin<sup>2</sup>
- Wanhu Sun<sup>2</sup>
- Yinyu Nie<sup>3</sup>
- Weikai Chen<sup>4</sup>
- Xiaoguang Han<sup>#1, 2</sup>
- 机构:
- SSE, CUHKSZ
- FNii, CUHKSZ
- Huawei Noah’s Ark Lab
- DCC Algorithm Research Center, Tencent Games
- 出版: ACM Transactions on Graphics (SIGGRAPH Asia 2024)
数据集资源
- ARXIV: https://arxiv.org/abs/2411.03047
- PDF: https://garverselod.github.io/GarVerseLOD.pdf
- CODE: https://github.com/zhongjinluo/GarVerseLOD
- DATA: https://github.com/zhongjinluo/GarVerseLOD
数据集摘要
- 摘要: 我们提出了一种分层框架,通过利用GarVerseLOD数据集中的服装形状和变形先验,来恢复不同层次的服装细节。给定从互联网搜索到的单张穿着人类图像,我们的方法能够生成高保真的3D独立服装网格,这些网格表现出逼真的变形,并与输入图像很好地对齐。
数据集详细描述
- 方法概述: 给定一张RGB图像,我们的方法首先估计T形服装形状,并借助预测的SMPL身体计算其与姿势相关的变形。然后使用像素对齐网络重建隐式精细服装,并采用几何感知边界估计器来预测服装边界。最后,我们进行服装注册以获得最终网格,该网格在拓扑上是一致的,并且具有开放边界。
- 数据集特点: GarVerseLOD收集了6000个高质量的服装模型,这些模型由专业艺术家手动创建,具有细粒度的几何细节。数据集被设计为具有不同细节层次的分层数据集,从无细节的样式化形状到与姿势混合的服装,具有像素对齐的细节。
- 评估: 我们的方法在大量野外图像上进行了评估,实验结果表明,GarVerseLOD能够生成比现有方法质量更高的独立服装部件,同时在姿势、光照、遮挡和变形的大量变化中表现出鲁棒性。
引用
bibtex @article{luo2024garverselod, title={GarVerseLOD: High-Fidelity 3D Garment Reconstruction from a Single In-the-Wild Image using a Dataset with Levels of Details}, author={Luo, Zhongjin and Liu, Haolin and Li, Chenghong and Du, Wanghao and Jin, Zirong and Nie, Yinyu and Chen, Weikai and Han, Xiaoguang}, journal={ACM Transactions on Graphics (TOG)}, year={2024} }




