多视图一致性约束的神经隐式表面重建  

Neural implicit surface reconstruction with multi-view consistency constraints

作　　者：魏文洁 娄路[1] Wei Wenjie;Lou Lu(School of Information Science and Engineering,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学信息科学与工程学院,重庆400074

出　　处：《中国图象图形学报》2025年第1期225-239,共15页Journal of Image and Graphics

基　　金：国家自然科学基金项目(82160345);重庆市自然科学基金项目(cstc2021jcyj-msxmX1121)。

摘　　要：目的通过体渲染学习的神经隐式表面表达是近两年涌现的多视图三维重建方法。针对具有开放表面非水密特点的目标物体,基于无符号距离函数(unsigned distance function,UDF)的神经隐式表面重建能够解决有符号距离函数(signed distance function,SDF)神经隐式表面重建方法的缺陷,重建精度更高。但是,该方法难以捕获多视图特征,对结构复杂、自遮挡的物体的三维重建容易出现模型缺失、形状偏差等问题。鉴于此,提出一种将多视图一致性约束与神经隐式表面重建相结合的方法。方法首先介绍在体渲染阶段使用的渲染权重和采样权重函数,确保实现无偏渲染,提高体渲染阶段精度以约束UDF网络的训练。然后,提出多视图几何一致性约束,关注重建模型中较大的光滑区域,通过显式约束获取更多复杂几何和纹理信息。最后结合多视图几何重建中提取特征的卷积网络,获取多视角的二维图像特征,通过最小化多视图深度特征提高学习隐式表面的神经网络的重建精度。结果实验在DTU数据集、DF3D(DeepFashion3D)服装数据集以及自采集的植物数据集上得到了更精细的表面,渲染结果更接近原始图像。在DTU数据集上与基准NeuS方法相比,倒角距离表示的几何结构误差平均值降低了0.16 mm,峰值信噪比表示的重建指标平均值提高了1.04 dB;在DF3D数据集上与NeuS、NeUDF方法相比,倒角距离平均值分别降低了2.76 mm、0.11 mm。重建得到的植物模型与服装模型具有高精度和完整的结构。结论本文提出的多视图一致性约束的神经隐式表面重建方法充分利用多视角图像的特征信息,提高了三维重建模型的质量,在复杂结构(如生长过程中的植物)和具有开放表面非水密的物体的三维重建上明显优于现有的先进方法。Objective Three-dimensional reconstruction is a critical technology in the field of computer vision,with pro⁃found implications across diverse domains such as medicine,engineering,and cultural heritage preservation.Early 3D reconstruction methods heavily relied on manual measurements and modeling,thus being complex and error prone.With the advent of digital photography,image-based 3D reconstruction became increasingly feasible.This research aims to elevate the quality of multi-view reconstruction.Traditional 3D reconstruction techniques such as stereo vision matching and laser scanning often demand precise camera poses and are highly contingent on the quality of feature matching.They exhibit limited effectiveness in scenarios involving weak textures,Lambertian surfaces,and slender objects.Meanwhile,deep learning-based 3D reconstruction methods primarily hinge on image depth information and point cloud data.How⁃ever,they grapple with challenges such as limited resolution due to memory constraints and complex topological structures.To improve the quality of reconstructions in smooth object regions and enhance model completeness,particularly for intri⁃cate non-closed curved surfaces,and to address issues such as missing model data and shape deviations,this study com⁃bines traditional multi-view consistency constraints with neural implicit surface reconstruction techniques to improve sur⁃face reconstruction accuracy and completeness.Method Our approach has three main facets.In the unsigned distance function(UDF)network training phase,we introduce multi-view geometric consistency constraints and feature consistency constraints.During the volume rendering process,we calculate color loss to guide the UDF network to learn scene informa⁃tion more closely aligned with reality.Neural radiance fields(NeRF),a novel object and scene representation method,implicitly model both geometry and appearance information.Employing a fully connected network(specifically,multilayer perceptron)and volume rendering techniques,Ne

关 键 词：三维重建 神经隐式表面 几何一致性 多视图特征一致性 网格模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]