Deep Learning Based 3D Object Reconstruction Using A View Planner

3D reconstruction is one of the hot topic in area of robotic and computer vision.Classic methods make it via calculating the change of camera pose with dense views,leading to the limitations such as the number of views,the noise of sensor and the condition of reflection.This may lose some important structures and even influence follow-up work such as shape analysis and interaction when not satisfying these limitations.This paper utilizes the deep learning technology to solve these problems.It not only researches the relations of view prediction and 3D structure prediction but also designs a unified deep learning model to reconstruct the 3D structure while planning the view sequence.With this model,it can improve the veracity of reconstruction and the efficiency of information gain.First,in order to explore the relation of view prediction and 3D structure prediction,this paper proposes a novel 3D object reconstruction model architecture based on deep le arning and view planning.Moreover,in order to validate that view prediction promotes t he information gain and the reconstruction quality,this paper designs a 3D volume recon struction network for multi-view image sequence,which integrates view prediction netwo rk module and 3D reconstruction network module to a unified architecture and obtains a satisfactory result compared with other methods.It validates that deep learning is able to predict the 3D structure with few information and view planning can promote the inform ation gain efficiency,leading to a better reconstruction result.Next,aimed to solve the limitations such as memory occupation,shape presentation and the loss of geometric feature,this paper designs a 3D point cloud reconstruction mo del for the combination of multi-view images and geometric feature.With combined sem antic feature and geometric feature from multi-view image sequence,the model plans the view sequence and predict the 3D feature at the same time.The point cloud,which is a general 3D data presentation,is used to save memory and computation,and make local d etails presented better.In the experiment,the model obtains a better detail presentation a nd reconstruction veracity.