Researches On Texture Mapping For Large-scale 3D Reconstruction

With the development of technologies such as drones and parallel computing,imagebased three-dimensional reconstruction has been successfully applied in many fields.Texture mapping is of great significance for improving the realism of 3D models,and is one of the key problems of 3D reconstruction.In order to optimize the performance of texture mapping in large-scale scenes,a highly parallelized view selection algorithm is proposed based on the convex optimization method,which improves the computational efficiency of texture mapping.Deformation information is used to suppress the generation of distorted textures,and the fast adaptive partition algorithm for 3D models further expands the practicality of texture mapping in large-scale scenes.The information of images and 3D models is combined to optimize the color consistency of texture charts,and the robustness of the algorithm is verified through experiments.In this thesis,the specific problems in texture mapping algorithm are studied as follows:First,the computational efficiency of existing texture mapping algorithms are limited by the view selection procedure.In this thesis,a new efficient convex optimization-based approach,i.e.the irregular graph based continuous max-flow algorithm is proposed.Potts model is introduced to mathematically formulate the key specific view selection problem,Then the continuous max-flow algorithm under irregular graph is propose,and in this work it is proved that,based on duality principle,the convex relaxation of original Potts model can be efficiently solved using this algorithm.Eventually the textures are recovered based on solution to the convex relaxation.The parallelism of the algorithm is proportional to the triangular faces of the 3D mesh model,making the algorithm easily implemented and accelerated upon a modern parallel platform.Experiments show that the proposed algorithm can yield high-precision texture mapping results,and significantly improve the computational efficiency of texture mapping.Second,to handle the poor visual quality of complex outdoor textured models,this thesis first incorporates texture deformation information into the view selection procedure.which significantly suppress the generation of distorted textures and improves visual quality..The 3D mesh models generated by 3D reconstruction pipeline usually have huge data amount and redundant faces.In this thesis,an adaptive mesh partition algorithm leveraging the normal of the mesh is proposed.The mesh can be divided into multiple submodels using this algorithm.Processing each sub-model individually can reduce memory consumption,which makes it possible for GPU to handle large-scale 3D scenes.Meanwhile,different sub-models can be processed using different GPUs,which further accelerating texture mapping.Third,while optimizing color consistency of complex 3D scenes,existing texture mapping approaches show poor robustness and tend to produce models with color shift.This thesis proposes a global-to-local joint optimization method to improve the color consistency of textured models.The 3D models are rendered with colors using the correspondence between 3D model and multiple views.Then each texture chart is locally corrected to enhance color consistency,which is guided by the colored 3Dvertices..With a global color constraint,the joint approach can achieve a robust color correction effect on models with many outliers.