单目视觉的深度与位姿联合预测网络  

A JOINT PREDICTION NETWORK OF DEPTH AND POSE IN MONOCULAR VISION

作　　者：贾瑞明[1] 李彤 刘圣杰 苗霞 王一丁[1] Jia Ruiming;Li Tong;Liu Shengjie;Miao Xia;Wang Yiding(School of Information Science and Technology,North China University of Technology,Beijing 100144,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《计算机应用与软件》2021年第12期155-160,186,共7页Computer Applications and Software

基　　金：国家自然科学基金面上项目(61673021);北方工业大学学生科技活动项目。

摘　　要：深度图与相机位姿参数是图像三维场景重建的重要数据,使用两个卷积网络分别预测,不仅效率低并且切断了二者之间的联系。对此提出一种联合预测深度图与相机位姿的卷积神经网络,输入单幅RGB图像,经过共享编码器编码,经两路子网络分别解码输出深度图与相机位姿参数,其中位姿预测子网络也为双路结构,将位置与姿态参数分离,避免两类参数的串扰。该网络的多任务结构通过信息共享可提升预测精度和效率。实验验证了该方法的可行性与优异性。The camera pose and depth map are both important data for 3D reconstruction of the scene.Using two independent networks to predict separately will not only increase the complexity of the algorithm,but also cut off the intrinsic link between the pose and depth maps.Aimed at this situation,a joint CNN combined with the camera pose and depth map is proposed.The input RGB image was encoded by the shared encoder,and then decoded by two sub-networks to output the pose and the depth map.The pose sub-network was also a dual-stream structure,which handles the position and rotation separately,avoiding the interference between two types of parameters.This multi-task structure can improve the prediction accuracy and efficiency of each task.The feasibility and superiority of the method are verified by experiments.

关 键 词：卷积神经网络 深度图预测 相机位姿估计 多任务结构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]