Intelligent Evolution Of Scene Space Layout For Large-Scale 3D Scene Modeling

The construction of 3D models for large-scale scenes is an important task in application areas such as computer games and digital twins.The large-scale scene modeling task,which is to plan and guide the placement of 3D objects under certain spatial area constraints,needs to provide an efficient solution from 2D scene layout planning.In this thesis,the intelligent evolution of scene space layout for large-scale 3D scene modeling is studied,and a total of three subdivision methods are proposed,the intelligent evolution of scene space based on WFC(Wave Function Collapse)and deep networks,the adaptive planning of scene space under boundary constraints,and the development of a scene space layout intelligent generation system to assist the design workflow.(1)A scene space intelligent evolution algorithm based on wave function collapse algorithm and deep network is proposed for the two-dimensional scene layout planning problem under soft boundary constraints.The algorithm consists of three modules:a wave function collapse algorithm for large-scale generation of initial layout wireframe patterns,a deep neural network for delineating reasonable boundaries of layouts from each layout wireframe graph,and another deep neural network for recommending functional semantic labels of subspaces.To meet the demand of multi-scale 3D scene diversity layout,a semantic phantom is also used for semantic mapping and variation to generate multi-scale and diversity layouts.A knowledge base containing 10,000 scenespace layout maps was constructed based on this algorithm.(2)A parameterized adaptive scene space planning algorithm based on differentiable rendering is proposed for the 2D scene layout planning problem under hard boundary constraints.The algorithm consists of three main parts:first,the constraints such as boundary constraint contours,initial mesh primitive groups and primitive deformation weights are input as adjustable parameters;second,the initial primitives are computed iteratively by deformation fitting based on the differentiable rendering optimizer jointly with three sets of constraint penalty terms;finally,the computed mesh primitives are optimized by post-processing,flattening the boundaries and recommending door and window positions according to the connectivity relationship,and finally the The result of the algorithm is converted into an indoor floor plan.The algorithm generates the layout plan adaptively without relying on a large amount of data training.(3)For the application goal of assisting two-dimensional scene layout planning,an intelligent generation system of scene space layout combined with intelligent algorithm is developed and implemented.The system is a web-side application,including three main parts:constraint input panel,result retrieval directory and layout editing panel.The user enters the boundary and constraint configuration in the input panel,and the system retrieves the layout of the data set similar to the input configuration and displays it in the retrieval catalog,where the layout is selected for hard or soft fitting to the input boundary,and finally the user can edit the layout configuration in the edit panel,and also subdivide each subspace of the layout recursively to heuristically assist the 2D layout design process and export the 3D modeling guidance data.To prove the effectiveness and usability of the methods in this thesis,intensive experiments were conducted for each method design,including quantitative index evaluation,qualitative analysis experiments,ablation experiments,control experiments,and comparison experiments with state-of-the-art methods,etc.It was verified that the three methods proposed in this thesis can effectively plan indoor spatial layout adaptively,generate heuristic to adapt to multi-scale scene layout,and assist scene layout planning through interactive systems to achieve the goal of large-scale 3D scene modeling.