Research On Technology Of Example-based Modeling With Structure And Geometry Generation

With the development of 3D scanning,rendering and fabricating technology,the 3D digital model has been applied widely among numerous fields,such as entertainment,medical treatment and industrial manufacturing.How to obtain new 3D models in an efficient way is a fundamental and significant problem in computer graphics.As a mainstream method,modeling by examples has received wide attention since introduced.Modeling by examples starts from the existing 3D models,finds out the rules of 3D models' construction,and guide the 3D modeling by these rules to simplify the modeling process.As to modeling by examples,how to enhance the reasonability and the variety of the generated models is a key problem.In this thesis,we focus on different grading of models construction rules.The contribution of this thesis including:(1)We design a structure-geometry separately framework for modeling by examples.In this framework,the construction rules of 3D models are divided into the globally structural priori knowledge and the locally geometrical priori knowledge.We first apply the segmentation and corresponding for 3D models to separate these two kinds of priorities.Then the representation learning is employed to gain the patterns of the global structure and the local geometry.Finally these two kinds of information are combined to construct a new 3D model.This framework take different grading of composition information to enhance the variety of the generated models.(2)We propose and implement a model structure generation method based on representation learning.To handle the diversity of models' structure,an isomorphic representation is presented.On the basis of the isomorphic representation,we introduce the full-connection AutoEncoder to learn the distribution of the isomorphic representation,and sample in the encoding space to generate new model structure.In addition,we design two validity constraint to detect the generated structure and visualize the detection result.(3)We design and implement a part geometry generation method based on representation learning.In this method,the geometry of model parts is represented as voxel.The 3D-convoution AutoEncoder is employed to learn the rule of formation for each type of model parts and new geometry of parts is produced by sampling in the encoding space.Moreover,we present a size normalization strategy to eliminate the influence caused by diverse parts' scale.The adversarial learning is also introduced to enhance the plausibility of the generated results.