Creative Modeling Of 3D Character Based On Semantic Tree

It is well knowed that one of the foundations of the virtual world, 3D model is an important object for digital design in the VR Era. Generally, the core aim of computer aided modeling research is to make the designers work more efficiently, and pursue more excellent designing results. In order to help the designers, the research of 3D modeling has gone through the stage of parameterization, characteristic and structure-aware, which was a process from concrete to abstract, from local to global,and from geometry to semantics. Until now, many methods have been proposed for model generating and editing. In this field, how to better support and guid creative modeling is a new topic in recent years. Learning by example is one of the main methods to solve this problem. For example, there are some creative modeling works based on part synthesis, the core ideas of which is batch learning and replacement in the same position. But these ideas are not applicable to the animation character in games or movies because they have variable structures and be full of imagination, which bring the difficult of data collection and the lack of structural innovation. With the design ideas becoming more and more personalized, the design of artificial objects will pursuing variable structures like the characters. Therefore,a representation of 3D model for variable structures and supporting structural innovation should be explored while the corresponding method for solving synthesis design should be founded.In this thesis, the guidance rules and effective methods of batch segmentation/labeling and synthesis modeling are discussed in the case of 3D characters in games as design objects. A representation of 3D model based semantic tree is proposed while a complete process of character innovation through segmentation and synthesis is created. The work includes the following four aspects.(1) An effective description of semantic tree is proposed in order to represent the structure of 3D character model. After analogy analysis of the monsters in game and the vertebrate species in nature, symbolic description and base view of the semantic tree for the character is designed. Then branch view, skeleton view and design view are constructed to support the ideas of segmentation and synthesis. Finally, aimed to the problem of the 3D character being less and monotonous in public database, a small character database is built, with data collected from diffrent sources.And, the views creation of characters in this database are tested and analyzed.(2) A method of batch segmentation and labeling for 3D character model with the same structure is proposed. Guided by the branch view, the constraint rules for the segmentation sequence and segmentation feature of different branches are designed.First, the ends of branches are confirmed by geodesic distance in proper order. Second,the segmentation vertexes are founded through searching the significant region of curvature variation. Finally, the shortest closed loop over the vertex is selected as the cut of segmentation in order to meet the need of part synthesis. This method can adapt rapidly to the segmentation mission of character model with different structure, which can solve the problems that 3D characters have variable branch structures and are difficult to be collected as training data.(3) A method of segmentation and labeling based on skin information for complex 3D character model is proposed. The skin is a kind of correlative information between vertex and skeleton for animation character, which reflects the similar movement trend of the vertices from the same part. So the skin would give the initial segmentation result by semantic annotation of the skeleton view before the cut is topological modified and optimized. This method can deal with the models that aren’t segmented by using existing methods to get meaningful result. These models have complex structure or visual interference or non-watertight in general.(4) A method of character model co-evolution is proposed based on replacement in the different joint. Guided by the design view, the evolution is set up to replace the child-tree gene group in a team of character. The probability graph is used to filter the unreasonable synthesis results rapidly after learning the structure. Further more, the steady search in solution space is maintained by using cooperative strategies that include keeping initial generation, the result from multiple characters having priority,the result from multiclass parts having priority, etc. This method can create new characters that can not be produced by replacement in the same joint. Besides its visual effect, creativity and inspiration have all been affirmed by the users.