Unsupervised Learning Based 3d Scene Reconstruction

The problem of three-dimensional scene reconstruction is to restore the three-dimensional scene structure from the two-dimensional image through equipment or deep learning.This problem is also a hot research issue in computer vision.At the same time,three-dimensional reconstruction technology plays an important role in many scenarios such as environment perception,robot navigation,scene interaction,three-dimensional digital protection of ancient buildings and so on.With the deeply studying of deep learning technology,especially deep neural network technology,and the disclosure of large-scale three-dimensional datasets,new possibilities and goals have been injected into the solution of three-dimensional scene reconstruction problems.To this end,this article will focus on the problem of three-dimensional scene reconstruction,using deep learning tools to carry out related research and exploration.First,this article will introduce a model framework for three-dimensional scene reconstruction based on unsupervised learning.The model extracts feature information from two-dimensional video images,uses unsupervised learning to predict the depth information and corresponding camera pose information for each frame of the video images,and restores the three-dimensional scene through depth information and camera pose information,then fuses and calculates to obtain the three-dimensional scene color point cloud model.After,the scene point cloud is reprojected to a two-dimensional plane according to the predicted camera pose,and then compared with the original video image,the difference is computed to optimize the model.The method proposed in this paper can effectively avoid the dependence of deep learning methods on real scene supervision information during the reconstruction process,and at the same time,it can establish a high-quality point cloud scene model.Later,for the study of this problem,considering that when using another different types of dataset in deep learning,the model build needs to be retrained to adapt the new dataset.Thus this article decomposes the previous research content and attempts to use a new meta-learning idea to predict the depth map information.It is expected that after multi-task training,when the network receives new depth map data,it can promote the current learning process according to the previously learned empirical knowledge.This process first learns a depth map prediction meta-model using different depth map data sets.Then,when new depth map data is added,it is possible to obtain the depth map information as accurate as possible through few iterations of training.Finally,this paper summarizes the proposed three-dimensional scene reconstruction method and the meta-learning depth map prediction model framework,analyzes the advantages and limitations of the method,and looks forward to the future research direction.