Research On Single-View 3D Reconstruction Based On Synthesis Multi-View

3D reconstruction aims to recover the 3D spatial structure based on 2D images of objects.There are rapidly growing interests across various industries for 3D reconstruction techniques,from computer visualization to auxiliary geometric design.Although the 3D reconstruction approaches based on depth map or RGB images have made some progress,while they generally suffer from several limitations: 1)The 2D to 3D learning strategy is brutal-force;2)These works ignore the effect of differences in appearance from different viewpoints of objects;3)Multiple images from distinctly different viewpoints are required.To solve the above problems,this thesis proposes a single-view 3D reconstruction algorithm based on synthesized multi-view.It aims to construct deep generative model synthesize multi-view based on single-view,incrementally improve the 3D reconstruction details of objects.(1)For the problem that the single-view is affected by the observation viewpoint and loses a lot of spatial information of the object,a single-view 3D reconstruction algorithm based on structure-aware are proposed.It learns to map the distributions between different-view 2D images and eventually generate multi-view images from a single image.The generated multi-view images are then employed to synthesize 3D shapes with incremental object details.In addition,by building perfectly aligned object poses in cross-view images as well as the corresponding 3D shape,the 3D reconstruction network can be guided to be aware of spatial structure of the objects.Finally,extensive quantitative and qualitative results demonstrate that the proposed approach improves the average IoU by2.97% on the ShapeNet dataset.(2)For the problem that the single-view 3D reconstruction algorithm based on structure-aware has rigorous requirements on the input data,an end-to-end view-aware 3D reconstruction network(VA3D)are proposed.The VA3 D includes a multi-neighbor-view synthesis sub-network and a 3D reconstruction sub-network.The multi-neighbor-view synthesis sub-network generates multiple neighboring viewpoint images based on the object source view,while the adaptive fusional module is added to resolve the image blurry and edge distortion issues in viewpoint translation.The 3D reconstruction sub-network introduces arecurrent neural network to recover the object 3D shape from multi-view sequence.Experimental results show that this algorithm effectively improves the 3D reconstruction results based on single-view,and a significantly lower variance in terms of reconstruction results compared to the single-view based approaches.