首页|基于隐式表达的服装三维重建

基于隐式表达的服装三维重建

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
随着近年来互联网购物的快速发展，在各大平台上出现了越来越多的服装商品。通过三维重建技术生成服装的三维模型可以帮助消费者更好地了解服装的姿态信息。针对服装的三维重建技术进行研究，提出基于隐式表达的服装三维重建模型。使用神经网络学习获得的占用函数作为服装三维模型的隐式表达，从而建立三维坐标和模型形状的映射。目前已有的三维重建算法需要拟合复杂的曲面模型，但资源消耗量大，而基于隐式表达的三维重建算法不需要进行参数化和网格化，能够提高算法的运行速度。为了进一步提高三维重建效果，采用目前性能最好的PointMetaBase-L网络模型和偏移注意力模块作为模型的特征提取网络。其中PointMetaBase-L网络模型基于现有的点云特征提取网络提出Set Abstraction层的元架构PointMeta，并通过分析选择PointMeta元架构中4个模块的最佳实践构成PointMetaBase-L网络模型的Set Abstraction层，同时引入平面特征投影模块加强特征的局部信息。在特征解码阶段，利用特征权重网络通过加权平均算法获取三维空间中采样点的占用概率。根据这些采样点的占用概率，通过基于区域增长的Marching Cubes算法提取高精度网格重建模型。实验结果表明，与占用网络相比，改进模型在交并比、倒角距离、法线一致性和F1值上分别提升了 48。83％，55。17％、4。27％和79。10％。
Implicit-Expression-based 3D Reconstruction of Clothing
Owing to the rapid development of Internet shopping in recent years,clothing items have appeared increasingly on major platforms.Generating Three-Dimensional(3D)models of garments using 3D reconstruction technology allows consumers to better understand the gestural information of garments.This study examines the 3D reconstruction technology of clothing and proposes a 3D clothing-reconstruction model based on implicit expressions.An occupancy function obtained via neural-network learning is used as the implicit expression of the garment 3D model,and mapping is performed between the 3D coordinates and model shape.Existing 3D reconstruction methods must fit complex surface models,which consumes a significant amount of resource,whereas implicit-expression-based 3D reconstruction algorithms require neither parameterization nor meshing,which accelerates the operation of the algorithm.To further improve the 3D reconstruction effect,the current best-performing PointMetaBase-L network model and the offset attention module are used as the feature-extraction network of the model.Between them,the PointMeta-Base-L network model proposes the PointMeta meta-architecture of the Set Abstraction layer based on the existing point cloud feature-extraction network and selects the best practices of four modules in the PointMeta meta-architecture to constitute the set of the PointMetaBase-L network model.This is performed by analyzing the abstraction layer while introducing a planar feature-projection module to enhance the local information of the features.In the feature-decoding stage,the occupancy probabilities of sampled points in 3D space are obtained using a weighted averaging algorithm via a feature-weighting network.A high-precision mesh-reconstruction model is extracted from the occupancy probabilities of the sampled points using the March Cubes algorithm based on regional growth.Experimental results show that compared with the occupancy network,the improved model improves the intersection ratio,chamfer distance,normal consistency,and F1 value by 48.83％,55.17％,4.27％,and 79.10％,respectively.

implicit expression3D reconstructionPointMetaBase-L network modeloffset attentionfeature weight networkregion growthMarching Cubes algorithm

费煜哲、蔡欣、赵鸣博、杨圣豪

展开 >

东华大学信息科学与技术学院,上海 201620

隐式表达 三维重建 PointMetaBase-L网络模型 偏移注意力 特征权重网络 区域增长 Marching Cubes算法

国家自然科学基金面上项目

61971121

2024

10.19678/j.issn.1000-3428.0067724

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(5)
  • 37