Research On Single Image 3D Point Cloud Reconstruction Based On Deep Learning

3D reconstruction is a long-standing key issue in the fields of computer vision and deep learning,and it has received extensive attention and research in recent decades.In the 3D reconstruction based on images,this task becomes a highly ill-posed problem due to the occluded part of the object.Although the use of images from multiple angles will reduce the difficulty of 3D reconstruction,in real life,not only is it easier to obtain images of a single angle of an object,but it is more practical to reconstruct the 3D shape from a single image.So we use a single image to complete the 3D reconstruction.However,since the self-occlusion part of a single image leads to an increase in unknown information,and the use of deep learning methods does not use accurate cameras to measure and photograph the target object,the existing methods generally suffer from low reconstruction quality of the occluded part of the object.The resulting shape has no precise details or even deviates from the actual problem.Aiming at the highly ill-posed problem of the3 D reconstruction of a single image,a method of gradually recovering the 3D point cloud of the object using the 2.5D sketch as an intermediary is proposed.This method first trains a 3D point cloud generator to learn the prior knowledge of the 3D point cloud in the data set.Then trained a model to recover the 3D shape from the 2D image.It first estimated the 2.5D sketch of the input 2D image,and then transferred the knowledge learned in the 3D point cloud domain to the 2D image domain through the 2.5D sketch.Aiming at the uncertainty of the occluded part of the target and the lack of precise details or even deviation from reality,a method of combining the deep generative model with the generative adversarial network is proposed.The method first uses the Gaussian probability distribution learned from the feature vector to predict the point cloud.Then the predicted point cloud is input to the discriminator in the generative adversarial network.When the output of the model is unrealistic,the point cloud generation network will be punished by the discriminator,so as to avoid the problem of the output point cloud model deviating from the actual situation.The proposed models are all tested on different data sets.The experiments show that this method effectively completes the high-quality 3D point cloud reconstruction,which is very competitive compared with the existing methods.