| Multi-view Stereo(MVS)vision estimation of dense representations of overlapping images is one of the central problems in computer vision research.In the past few years,efforts have been made to improve the quality of MVS 3D reconstruction tasks,and some traditional geometric algorithms,such as PMVS,GIPUMA,COLMAP,have achieved impressive results.However,traditional geometry-based multi-view 3D reconstruction algorithms have high accuracy but their integrity in weak texture regions is insufficient.The rise of deep learning has made image-based MVS reconstruction tasks possible,and existing deep learning-based methods for multi-view stereo(MVS)have achieved satisfactory results,but because these methods introduce a large ambiguity in low-texture or even no-texture or highly reflective regions,leading to inaccurate depth predictions,which directly compromises the integrity of the reconstruction results.In particular,most methods based on depth convolutional networks are dedicated to accuracy improvement,which leads to poor reconstruction in low-texture smoothed regions and thus loss of integrity.Based on the above problems,two different solutions are proposed in this paper.Compared with the existing methods,the proposed methods in this paper both have significant improvement in the completeness of MVS reconstruction and have certain robustness.Specifically.This paper first proposes a global balanced velocity estimation network for maximizing the overall completeness of the reconstruction results with guaranteed accuracy,which overcomes the problem of depth ambiguity in low-texture regions for the MVS task.Secondly,this paper proposes a depth estimation network based on pyramidal loss strategy,which obtains finer depth maps through multi-round iterative optimization and multi-round loss constraints.Finally,this paper designs and develops a multi-view real-time 3D reconstruction system.In summary,the contributions of this paper are mainly as follows:1.In this paper,a global balanced depth estimation network(GBNet)is proposed to solve the depth ambiguity problem of MVS in low-texture smoothing regions.The method uses a feature distillation module and a group attention module to refine high-resolution semantic features and enhance the sensitivity and learning ability of the network to global features.In addition,the method performs regularization by a non-standard 3DCNN network to aggregate the depth information to obtain a more complete depth map and finally a more complete reconstruction result.2.In this paper,a two-stage capsule depth estimation network(CapNet)is proposed,which uses a 2D capsule network to model local details in the first stage and a 3D capsule UNet network to fully encode high-dimensional features to model long-range overall dependencies in the second stage,this approach results in a more detailed and complete depth map,and the edge contours are distinguished from the background to obtain fine edge information,and finally achieve a high integrity reconstruction result.3.In this paper,a multi-view real-time 3D reconstruction system is designed and developed to realize the real-time 3D reconstruction process by calling 3D reconstruction algorithms,and then the vivid point cloud results are displayed on the web page. |
