基于堆叠循环沙漏网络的多视角立体重建  

Stacked recurrent hourglass network based depth inference for multi-view stereo

作　　者：刘晓玉 刘韵婷 郭辉[2] LIU Xiaoyu;LIU Yunting;GUO Hui(Shenyang Ligong University,Shenyang 110023,China;Shenyang University of Technology,Shenyang 110870,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110023 [2]沈阳工业大学,辽宁沈阳110023

出　　处：《通信与信息技术》2023年第3期37-42,共6页Communication & Information Technology

基　　金：辽宁省教育厅(项目编号:LJKMZ20220617)资助的课题。

摘　　要：针对多视角三维重建点云图精度低、边界模糊等问题,提出了一种基于堆叠循环沙漏结构的多视角深度估计算法。首先,网络使用2D U-Net进行特征提取,利用由粗到细的方式构建一个代价体金字塔,只在金字塔的第一层使用3DCNN卷积进行代价体正则化,其他层使用一个新式的堆栈沙漏模块,该模块由不同尺寸的卷积层和GRU层组成,然后通过两个不同尺寸反卷积层进行解码操作,最后回归图像的深度。实验在DTU数据集上与最新的10多种方法进行了比较,相比于性能第2的网络,精度提高了4.72%,平均整体性能提高3.98%。实现了对图像深度的准确估计,提高了深度图表面的清晰度和深度预测精度。在DTU数据集上进行了数值实验和主观评测,结果表明该算法获得了更好的三维重建效果。Aiming at the problems of low accuracy and fuzzy boundary of multi-view 3D reconstructed point cloud image,a multi-view depth estimation algorithm based on stacked cyclic hourglass structure was proposed.Firstly,the network uses 2D U-Net for feature extraction,and constructs a cost body pyramid in a coarse-to-fine way.Only 3DCNN convolution is used for cost body regu⁃larization at the first layer of the pyramid,and a new stack hourglass module is used for other layers,which is composed of convolution layers and GRU layers of different sizes.Then the decoding operation is carried out through two deconvolution layers of different sizes,and finally the depth of the image is returned.The experiment is compared with the latest 10 methods on the DTU data set.Compared with the second performance network,the accuracy is improved by 4.72%and the average overall performance is improved by 3.98%.The accuracy of image depth estimation is realized,the surface ambiguity of depth map is reduced,and the accuracy of depth prediction is improved.Numerical experiments and subjective evaluation on DTU data sets show that the proposed algorithm achieves better 3D reconstruction results.

关 键 词：多视角 三维重建 深度估计 点云 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]