With the rapid development of deep learning,research and application of 3D scene restoration based on multi-view geometrybecomes more and more popular. Many researchers focus?on improving the recovery effect of 3D scene by optimizing deep networkstogether with training datasets of regularly distributed camera poses. However,in practical application,when?
ordinary users shoot targetscenes,the camera pose distributions are highly random. It is difficult to ensure that the quality of obtained images be equal or close tothat of the training datasets,thus affecting the recovery effect. In order to alleviate this problem,a 3D scene restoration method based onweak random camera pose images is proposed. By providing users?
with suggestions for shooting target scenes,the randomness of camerapose distribution of the target scene is greatly reduced. Firstly,the user obtains 2D image data from different perspectives in the samescene under the guidance of target scene shooting suggestions,and then recovers 3D sparse point cloud and camera pose of the scenethrough Structure From Motion ( SFM) . Finally,the dense reconstruction of 3D point clouds is carried out in Multi-View Stereo ( MVS)network model. The experimental results show that the proposed method effectively reduces the random of camera pose distribution of theacquired target scene images compared with that before shooting suggestions,and the success rate of 3D scene recovery?
is increased by52. 95% .