基于学习的鲁棒三维射影重建  

Robust 3D Projective Reconstruction by Learning

作　　者：舒茂 胡立华[1,3] 董秋雷[1,2,4] 许华荣 胡占义[1,2,4] Shu Mao;Hu Lihua;Dong Qiulei;Xu Huarong;Hu Zhanyi(National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190;Department of Computer Science and Technology,Xiamen University of Technology,Xiamen 361024;University of Chinese Academy of Sciences,Beijing 100049;CAS Center for Excellence in Brain Science and Intelligence Technology,Beijing 100190;School of Computer Science&Technology,Taiyuan University of Science and Technology,Taiyuan 030024)

机构地区：[1]中国科学院自动化研究所模式识别国家重点实验室,北京100190 [2]中国科学院大学,北京100049 [3]太原科技大学计算机科学与技术学院,太原030024 [4]中国科学院脑科学与智能技术卓越创新中心,北京100190 [5]厦门理工学院计算机科学与技术系,厦门361024

出　　处：《计算机辅助设计与图形学学报》2018年第2期309-317,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61333015,61402316,61375042);太原科技大学校博士启动基金(20162009).

摘　　要：基于图像的三维重建是计算机视觉领域中一个重要的研究主题.针对目前深度神经网络无法有效剔除多幅图像对应点中的外点的问题,提出一种鲁棒的深度卷积神经网络,用以从多幅图像对应点中准确地恢复场景的三维射影结构.该网络首先把输入的对应点分为多个不同的子集,每个子集独立地进行射影重建;然后通过权重计算层得到每个射影重建的权重;最后通过合并层对这些不同的射影重建加权求和,得到最终的鲁棒的射影重建.实验结果表明,该网络具有较高的重建精度和很强的鲁棒性.Image-based 3D reconstruction is an important research topic in computer vision.The current deep neural networks cannot effectively eliminate outliers from point correspondences across multiple images.To address this problem,a robust deep convolutional neural network is proposed to accurately recover the 3D projective structure of scenes from point correspondences across multiple images.First,the network divides the input point correspondences into several different subsets,and each subset acts independently for a projective reconstruction;then,the weight of each projective reconstruction is estimated through a weight-learning layer;finally,a merging layer is activated to perform weighted summation of these different projective reconstructions to get the final robust projective reconstruction.Experimental results demonstrate both the reconstruction accuracy and strong robustness of our network.

关 键 词：射影重建 卷积神经网络 外点剔除 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]