Depth Estimation And Accuracy Evaluation Of Single Image For 3D Reconstruction Of Ancient Buildings

Chinese Buddhist and Taoist temples are mainly distributed on high mountains,and are an important part of ancient architecture in China.Most of these ancient buildings are wooden structures,which are vulnerable to erosion and damage from natural disasters,resulting in irreparable losses.As an effective means of protecting ancient buildings,3D digitization of ancient buildings has important theoretical significance and practical application value.The existing three-dimensional digital protection technology of ancient buildings mainly includes the method based on laser scanning and the method based on image three-dimensional reconstruction.However,affected by the environment of ancient buildings,the method based on laser scanning can not collect all the point clouds of ancient buildings.Therefore,this paper takes the local structure of ancient buildings as the object,and mainly studies the depth estimation and accuracy evaluation methods of a single image under the unsupervised framework.The main work is as follows:1.In order to effectively evaluate the accuracy of the depth estimation method of ancient buildings,binocular cameras and depth cameras are used to capture the local structure information of ancient buildings in 15 scenes in the Beijing Summer Palace scenic area,including beams,fangs,rafters and other unique structural parts of ancient buildings.A dataset of local structure images of ancient buildings is constructed.Among them,there are 4800 pairs of local scene images of ancient buildings captured by binocular cameras,and a number of local images of ancient buildings and corresponding depths captured by a depth camera.2.Aiming at the problem of low depth estimation accuracy by scale ambiguity in single-view depth estimation methods in outdoor and indoor scenes,combined with binocular vision theory and deep network,a single-image-based depth estimation network under unsupervised framework is designed.Based on the calibrated camera pose to estimate the depth,the network optimizes the scale of the baseline parameters in the input camera pose.Finally,the optimized network output is compared with the network output of the calibrated camera pose estimation depth of field,which can not only generate the image depth in the absolute scale,but also improve the generated absolute depth accuracy according to the existing evaluation indicators.In some scenarios,the absolute relative error Abs Rel,root mean square error RMSE and logarithmic root mean square error RMSElog are reduced by more than 30%.3.Aiming at the problem that the existing single image depth estimation methods cannot be applied to the evaluation of 3D reconstruction scenes,two evaluation methods for 3D reconstruction-oriented single image depth estimation are designed by calculating the absolute error of depth and the average distance of point clouds.Finally,the above two evaluation methods are applied to the experimental evaluation.The experimental results show that for the 3D reconstruction of ancient buildings,so far,the depth estimated by the depth network is still difficult to achieve the centimeter-level reconstruction accuracy required by the digital archive of ancient buildings,but the impact on the depth learning based on a single image is not obvious.4.Combined with the single image depth estimation method,a prototype system based on single image depth estimation is designed and developed.After inputting the dataset,the system independently selects the image of the ancient building structure to be tested,uses the trained network model to generate the depth of a single image,and displays the three-dimensional point cloud of the structure to the user at the same time.