3D Reconstruction From A Single-image Based On CNN Network

As a key technology in the field of computer vision,3D reconstruction is widely used in scene reconstruction,robot,unmanned driving,3D measurement,urban planning.Traditional 3D reconstruction technology based on multi-view and single-view often relies on professional equipment,and it is difficult to build accurate 3D model due to the problems of calibration,feature point matching and so on.With the rapid development of deep learning technology,3D reconstruction based on single image using data-driven method overcomes the problems such as matching.Based on three-dimensional point cloud data,this paper studies 3D reconstruction from a single-image based on CNN Network.The main contents are as follows:1.In order to solve the problem of low accuracy of point cloud in single image 3D reconstruction,a two-stage training deep learning network model 3DARNet is constructed.Firstly,the point cloud auto-encoder network is pretrained to make it have the ability to learn the features of point cloud data.Then the image encoder network with attention mechanism is used to combine with the pre-trained point cloud auto-encoder for the second training,and finally the whole network is formed.The network uses a single image as input,generates simple point cloud through image encoder,and then obtains accurate point cloud reconstruction effect through point cloud auto-encoder.2.In order to solve the problem that the number of point cloud generated from single image is limited,this paper establishes a 3D-UPCNet reconstruction model based on point cloud upsampling.A single image is used as the input of the image encoder,and the sparse point cloud is generated.The global feature,local feature and dynamic edge feature of the point cloud data are extracted in the point cloud upsampling module.Finally,the upsampled point cloud reconstruction is completed,and the reconstruction accuracy and visualization effect are improved.The two network models proposed in this paper are all tested on ShapeNet dataset and Pix3 D dataset.Through qualitative and quantitative analysis,the results of this paper have been improved compared with the state-of-the-art networks,which verifies that the two networks proposed in this paper can effectively improve the performance of 3D reconstruction from single image based on CNN network.