Analysis And Synthesis For 3D Indoor Scenes

In computer graphics, computer vision and other related fields, the modeling of various 3D scenes is always a research hotspot. With the rapid development of computer simulation, virtual reality and other scientific technologies, the scope of application of 3D indoor scene is also more extensive. Nowadays,3D indoor scene can be applied in some activities such as fire escape, hostage rescue, room escape, archaeological excavation and so on. Therefore, the modeling of 3D indoor scene is a great significant subject. Currently, the modeling methods of an object or scene can be divided into three types:using 3D modeling software, modeling by collecting information with instruments and equipment, modeling with image or video. Because there are some disadvantages of costing large time, complex to operate and bad modeling effect, the automated synthesis of 3D indoor scenes becomes a challenging problem in the relevant areas. This paper analyses the objects’geometry relationship and arrangement distribution, and based on which achieved automatic synthesis of 3D indoor scenes. The main contribution can be summarized as follows.(1) We propose three algorithms in order to detect and extract the relationship of intersection, proximity and supporting among objects. Besides, we also construct an occurrence network among objects who has supporting relationship, then we extract supporting planes of the input scene. The scene analysis provides theoretical basis for optimizing the scene synthesis algorithm.(2) We use Gaussian Mixture Model to train arrangement distribution. For each supporting plane, we first make the location of all objects parameterized and normalized. Then we store the data in a file. The account of the location of each object represents the frequency it occurs in the scene set. After that, the GMM is fitted to the locations of each category.(3) We propose a novelty algorithm which can synthesize 3D indoor scene automatically. The initial indoor scene that users specify and the description of objects to be inserted are taken as the input. Then, some objects are selected from the database, and automatically arranged on the appropriate supporters in the scene. The process of our method shows scene evolution of enriching simple scenes. Our method can not only satisfy the users’ design requirements, but also avoid the tedious manual operation.