Research On Monocular Vision 3D Point Cloud Reconstrucion Based On Deep Learning

It is a commonly used 3D modeling method to construct a 3D data model in a virtual 3D space through 3D production software such as CAD.However,when faced with a large number of complex modeling tasks,the time and labor costs are huge.While the automatic modeling methods based on binocular and trinocular vision solve the problems of huge manpower and time cost of complex modeling techniques,these methods are too dependent on the equipment.This paper researches and proposes a monocular vision 3D point cloud reconstruction technology based on deep learning,which can effectively reduce the cost of manual modeling,and also avoid the equipment cost and environmental limitations of using a depth camera.This paper first proposes a 3D-ReConstnet based 3D point cloud reconstruction network based on deep learning.The network uses residual convolutional neural network to extract the features of the picture,which avoids the problem of gradient exploding or gradient vanishing occurring in the depth neural network and the problem of difficult training of the depth neural network.And using Chamfer distance and Earth Mover's distance as the loss function to optimize the effect of the model.It can evaluate the distribution of point cloud correctly and avoid the problem of incorrect evaluation.At the same time,the two constraints of the two evaluation functions may also solve the problem that the reconstruction model is too decentralized or too centralized.The experimental results on ShapeNet and Pix3D CAD 3D model datasets prove that the proposed deep 3D point cloud reconstruction network has a good reconstruction effect.For occluded two-dimensional pictures,it is not suitable to give a single and certain three-dimensional reconstruction.Therefore,this paper proposes a method for generating semantic diversity reconstruction for two-dimensional pictures with occlusion.The proposed method first maps the feature vector of a two-dimensional picture into a probability vector of normal distribution,and generates a point cloud based on this probability vector.After that,the proposed method uses a diversity loss function based on the penalty angle to train a probability vector generation network to generate diversity reconstructions with different semantics.The proposed semantic diversity reconstruction method can generate diverse and reasonable reconstructions with different semantics for occluded pictures without affecting the reconstruction and generation of non-occluded pictures with sufficient information.