3D Shape Segmentation Based On Deep Mesh And Manifold Geometry

The number of 3D models is increasing explosively with the development of sensor technology and mechanical drawing.In the field of computer vision,some researchers have shifted their focus from 2D images to 3D models which better represent the real world.3D shape segmentation is the basis of 3D shape analysis,and it's also the basis of detecting and recognizing object information in computer vision.The shape of the 3D model is segmented to determine the semantics of each mesh in the model.Through deeper analysis of each semantic component,shape correspondence and matching,model retrieval and other tasks are realized.For the 3D shape segmentation of mesh,early scholars started from the perspective of computational geometry and calculated geometric properties of mesh to find the effective feature descriptors of the model in 3D space.In order to solve problems of single feature and non-robustness in earlier methods,3D shape segmentation methods based on deep learning are proposed.These methods are able to learn enough feature descriptions from the model,fully mine the geometric information contained in the shape,and automatically extract the features that are beneficial to the segmentation task.For improving the accuracy and reliability of 3D shape segmentation,the thesis designs two kinds of automatic 3D shape segmentation methods with high segmentation precision,based on deep mesh and manifold geometry.The main research contents are as follows:3D shape segmentation is obtained by projection image semantic segmentation.Optimized Dilated Convolution Network(ODCN)for 3D shape segmentation is proposed.This method firstly obtains projection images of the 3D shape.Then,the semantic segmentation of the projection images is obtained by using the ODCN.Later,the semantic labels are back projected to the 3D shape,a conditional random field method is used to optimize the segmentation result.Experiments on PSB dataset and COSEG dataset show that the ODCN method has higher label accuracy in shape segmentation task compared with other methods.To perform 3D shape segmentation by deep mesh,an Edge-face Feature Combined Dual Network(EFCDNET)for 3D mesh segmentation is proposed.The EFCDNET uses two network branches.The first branch uses the above-mentioned ODCN.The other branch designs the deep mesh based on edge features,according to triangular manifold characteristics of irregular meshes.The segmentation results of the two branches are fused by a voting algorithm,and then the final segmentation of 3D mesh is obtained.Experiments on COSEG dataset and HBS dataset show that the dual network can effectively improve the accuracy of 3D shape segmentation.Furthermore,neural network frameworks of Caffe and PyTorch are used to implement the ODCN and EFCDNET.A 3D shape segmentation program is written by PyQt5,which inputs 3D mesh models and outputs segmentation results on the mesh faces or edges.