Multi-view Stereo Reconstruction Based On Global Structure From Motion

3D reconstruction is one of the most popular research fields in computer vision.It has been widely used in animation,virtual reality,and biological image and mixed reality.With the reduction of the price of high-quality cameras and the improvement of computer computing power,the application and research of 3D reconstruction based on images is more and more extensive,and the corresponding 3D scenes can be restored according to any 2D image,which has broad prospects in the future.The global structure from motion can restore the poses of all cameras and the coordinates of three-dimensional points according to the images of any scene only one time,which is an efficient three-dimensional reconstruction method.Solving the camera poses is the core part of the global SFM,therefore,we propose a new global camera pose estimation algorithm WTLS-IRLS,which can effectively solve the global rotation when there are outlier.Firstly,according to the relationship between the rotation vector and the rotation matrix,we simplify the product operation of the rotation matrix into the subtraction operation in the rotation vector,which reduces the complexity of the algorithm.Secondly,weighted total least squares and the iteratively reweighted least squares are used to perform the robust average of relative rotation,and finally the optimized global rotation is obtained.After obtaining the global camera pose based on WTLS-IRLS algorithm,the sparse point cloud can be further obtained,but the sparse point cloud cannot meet the requirements of visualization.Therefore,we propose a quasi-dense matching algorithm between images based on the principle of matching propagation.The algorithm starts with a set of sparse seed matching,then propagates to neighboring pixels through the best priority strategy,and enforces gradient constraints and unique constraints,and finally obtains a dense disparity map,and restores dense three-dimensional point clouds.The experimental results show that WTLS-IRLS has an improvement in efficiency,accuracy and iterations.For the quasi-dense matching algorithm based on matching propagation,it is robust to the outliers of the initial sparse matching because of the best priority strategy,and the complexity in time and space is low.Based on experiments with multiple real-world datasets,we obtained high-quality dense matching results,and finally obtained dense and accurate 3D point clouds based on the results of quasi-dense matching.