A Multi-View Stereo Reconstruction Study Based On Costal Pyramids

Due to its advantages of low cost,quick data collecting,and high realism,3D reconstruction based on multi-view images has become a research hotspot in the field of computer vision.The cost volume pyramid is used to develop a multi-view stereo reconstruction technique in this paper.The system investigates and improves multi-view stereo reconstruction’s main phases.(1)Because the loop algorithm’s multi-view 3D reconstruction is a permutation version,the reconstructed 3D forms of different configurations of the same image may differ.As a result,this study includes an AttSets attention aggregation module at the feature extraction stage.The feature extraction method based on AttSets is permutationinvariant and can learn to fuse information from all images,resulting in improved aggregation performance.(2)After collecting the extracted features,this article considers the memory and time constraints of generating the cost volume and uses the cost volume pyramid structure to iteratively optimize the cost volume construction.The mistake of the initial depth map will be transferred to the depth map of the final layer,resulting in an unsatisfactory final reconstruction effect because the depth map is created in a residual iterative way.To refine the original depth map and improve the final depth map quality,depth normal vector consistency is introduced to the first layer of pyramid iteration in this paper.(3)In recent years,the loss function in multi-view stereo reconstruction models based on the residual technique has primarily used pixel-by-pixel loss.However,because elements such as illumination,picture translation,and so on induce pixel alterations,pixel loss will result in a big erroneous value.As a result,in order to optimize model training as much as feasible,this paper includes the feature loss in addition to the pixel loss.This research uses a DTU dataset created specifically for MVS reconstruction to test and train.In order to illustrate the algorithm’s generalization capabilities,it is also tested on tanks,temples,and self-selected datasets.Traditional and deep learning-based methodologies are compared to the test results.The experimental findings suggest that the technique presented in this research achieves the best completion results.The algorithm is capable of entirely reconstructing the scene,and the restored scene is of excellent quality.In addition,a series of pairs experiments were carried out to determine the ideal number of matches,the number of pyramidal layers in the costomer,and the algorithm’s pixel depth interval.The findings of the experiment reveal that the number of views,the number of likes,and the number of shares are all related to the number of views.