Learning-based Surface Reconstruction Methods For Large-scale Point Clouds

The reconstruction of surfaces from point clouds for better representation and manipulation is an important foundation for 3D modeling and its applications in 3D rendering,virtual reality,and augmented reality.The traditional geometric surface reconstruction methods have achieved great success and have been widely used in practice.However,they still face problems of low scalability and low reconstruction quality.With the development of deep learning,solving reconstruction problems using learning-based methods has raised great research interests.However,existing learning-based method is greatly inferior to traditional methods in terms of data scale and reconstruction quality,and their application is thus limited.Through analysis of the feature construction of the traditional geometric method and the deep learning frameworks,this paper integrates the core concepts of feature construction in geometric methods into the deep learning methods,extracting the geometric structure features directly related to surface.Meanwhile,network architectures and reconstruction pipelines are design to be adaptive to complex point clouds.The proposed methods focus on solving the problems of non-scalability,inability to reconstruct geometric details and poor generalization capability faced by learning-based methods,so that learning-based methods can be compared with traditional geometric methods in terms of reconstruction quality and reconstruction scale.This provides an important foundation for future application of learning-based methods.The research contents and innovations are summarized as follows:First,the basic connotation,input and output formats of surface reconstruction are introduced and analyzed in detail in order to provide background concepts.Then,the core concepts and common algorithms of geometric methods are also introduced in detail.The geometric feature construction and function fitting methods such as signed distance and normal direction in geometric methods are analyzed.Then the framework and problems of learning-based methods are analyzed,focusing on the analysis of their network architectures and feature construction problems.This provides ideas for integrating the core concepts of feature construction in geometric methods into deep learning networks.Second,a learning-based scalable surface reconstruction method for point clouds is proposed,which solves the scalability and quality problem of learning-based surface reconstruction methods for the first time.The reconstruction quality and generalization capability of learning-based method are greatly improved.This method classifies vertices of an octree as in front and at back of the implicit surface using networks independent to octree structure for scalability.A network layer for consistent surface feature extraction from point clouds is proposed.The local geometry-aware feature of octree vertices of the octree is designed in order to reconstruct the local geometric details and improve the reconstruction accuracy and generalization capability of the network.Experiments show that this method can reconstruct high-quality surfaces for large-scale point clouds at a high time efficiency,and requires very little training data to obtain high accuracy and generalization capability.Third,a learning-based reconstruction method from Delaunay triangulation of point cloud is proposed,which applies deep learning methods on Delaunay triangulation for the first time and obtain high-quality reconstruction results.The Delaunay triangulation is seen as a graph.The geometric features are aggregated into the graph in order to form a feature augmented graph model.A graph filtering module and a neighborhood consistency loss are designed and can accurately predict the in/out labels of the tetrahedrons in the Delaunay triangulation.The result surface is extracted from those labeled tetrahedrons.Experiments show that this method can reconstruct high-quality surfaces adaptively for complex point clouds,and it has good generalization capability on different types of data.