多阶段协同的多视图立体算法研究  

Research on multi-view stereo algorithm based on multi-stage cooperation

作　　者：朱治年 刘韵婷 肖培宇 李思维 刘欣然 ZHU Zhinian;LIU Yunting;XIAO Peiyu;LI Siwei;LIU Xinran(Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]沈阳理工大学自动化与电气工程学院,辽宁沈阳110159

出　　处：《通信与信息技术》2025年第2期28-32,共5页Communication & Information Technology

摘　　要：多视图立体匹配(Multi-View Stereo)是计算机视觉领域的重要任务之一,旨在从多个视角的图像中恢复场景的结构信息。然而,由于成本体积聚合在局部存在着严重的不一致性,直接聚合几何相邻成本会导致严重错误导向。现有的方法要么寻求二维空间的最优选择性聚集,要么增加聚集的手段,但都无法有效解决成本体积的几何不一致性,导致深度估计的精度和鲁棒性不佳。为了解决这个问题,提出用于多视图立体的协同表达(CRMVS),旨在协同多个模块整合几何的一致性信息,提高多视图立体匹配任务的深度估计精度和鲁棒性。首先,利用改进的特征金字塔网络(FPN)增强网络的特征提取能力。其次,设计了一个渐进式权重网络模块(PWN)进行代价体的构建。最后,设计了一个几何代价聚合与精化网络模块(GCR)来对代价体进行精准聚合。实验结果表明在DTU,Tanks&Temple数据集上都展现出了先进的性能。Multi-View Stereo is one of the important tasks in the field of computer vision,which aims to recover the structural infor⁃mation of a scene from images from multiple perspectives.However,due to the severe local inconsistencies in cost-volume aggregation,direct aggregation of geometrically adjacent costs can lead to serious misdirection.The existing methods either seek the optimal selective aggregation of two-dimensional space,or increase the means of aggregation,but they cannot effectively solve the geometric inconsistency of cost volume,resulting in poor accuracy and robustness of depth estimation.In order to solve this problem,a Collaborative Representa⁃tion for Multi-view Stereo(CRMVS)was proposed,which aimed to integrate the consistency information of the geometry with multiple modules and improve the depth estimation accuracy and robustness of the multi-view stereo matching task.Firstly,we use the improved Feature Pyramid Network(FPN)to enhance the feature extraction capability of the network.Secondly,we design a progressive weighted network module(PWN)to construct the cost body.Finally,we design a Geometric Cost Aggregation and Refinement Network Module(GCR)to accurately aggregate the cost body.Experimental results show that our method shows advanced performance on DTU,Tanks&Temple datasets.

关 键 词：多视图立体匹配 特征金字塔 代价体 成本聚合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]