Realistic Tree Modeling With Inverse And Intelligent Interaction Technologies

With the development and popularization of 3D games,entertainment and virtual reality applications,the requirements on diverse and compelling 3D models are increas-ing day by day.Trees are indispensable objects in natural scenes and realistic tree models can greatly enhance the authenticity of the virtual scene by providing users more immer-sive experience.Therefore,generating botanically-plausible tree models under certain shape constraints and designing interface for ordinary users to model trees in an intuitive and efficient way,have both been the hot topics in graphics research.These topics both have important research significance and great application prospects.At present,to generate botanically-plausible tree models,most of the studies fo-cused on simulating the tree's growth mechanism.These bottom-up methods need the user to explicitly specify certain botanical rules in order to generate tree models with desired patterns.By lacking the efficient control of the whole tree shape,these locally adaptive methods are not only limited in the capability of modeling trees with specified shapes but also in supporting ordinary users modeling trees intuitively and efficiently.Thus,researchers have paid attention on the inverse methods for tree modeling.But the inverse modeling problem is challenging due to its ill-posed nature.The current inverse methods not only have a variety of restrictions on the input but also time-consuming.Besides,these methods only pay attention to the design of specific modeling algorithms but ignoring the also very important fact on designing a user-friendly interface.In this thesis,we further studied the aforementioned problems and proposed a new kind of in-verse tree modeling algorithms and a new intelligent modeling interface designed for ordinary users.Our main contributions are as following:By considering a number of botanical rules,a botanical representation of trees has been presented to provide rich morphological controls with a small number of parameters.Built upon it,a further semantic representation has been presented through crowdsourcing data.These representations provided an important compu-tation basis for our inverse algorithms and intelligent interaction.Based on the botanical representation,a variational inverse modeling algorithm has been presented.The inverse tree modeling problem has been formulated as a variational optimization problem,where the shape and other botanical constraints could be unified into one mathematical framework.An efficient optimization algo-rithm considering the branching structures of trees has also been presented,which is one to two magnitudes faster than existing work.Based on the tree's semantic representation and our variational inverse modeling algorithm,a user-intent guided and intelligent tree modeling interface has been presented.By combining the Kalman filtering with the upper confidence bound skillfully,an online learning algorithm has been designed to learn a user intent evaluation function effectively through limited interaction data,which then sup-ported the use modeling trees intuitively by exploring the underlying semantic trait space of trees.A framework for single image tree modeling has further been presented based on the above work.By combining the deep learning with our variational inverse modeling algorithm,this framework can model trees according to real tree photos.A data set containing a large number of 3D tree models and rendered images has been created to train the deep neural network.The finally trained network can not only support scalable modeling but also solving the depth problem caused by varying degrees of self-occlusions in a consistent way.